NDArray API

Overview

This document lists the routines of the n-dimensional array package:

mxnet.ndarray NDArray API of MXNet.

The NDArray API, defined in the ndarray (or simply nd) package, provides imperative tensor operations on CPU/GPU. An NDArray represents a multi-dimensional, fixed-size homogenous array.

>>> x = mx.nd.array([[1, 2, 3], [4, 5, 6]])
>>> type(x)

>>> x.shape
(2, 3)
>>> y = x + mx.nd.ones(x.shape)*3
>>> print(y.asnumpy())
[[ 4.  5.  6.]
 [ 7.  8.  9.]]
>>> z = y.as_in_context(mx.gpu(0))
>>> print(z)

A detailed tutorial is available at NDArray - Imperative tensor operations on CPU/GPU.

Note

mxnet.ndarray is similar to numpy.ndarray in some aspects. But the differences are not negligible. For instance:

  • mxnet.ndarray.NDArray.T does real data transpose to return new a copied array, instead of returning a view of the input array.
  • mxnet.ndarray.dot performs dot product between the last axis of the first input array and the first axis of the second input, while numpy.dot uses the second last axis of the input array.

In addition, mxnet.ndarray.NDArray supports GPU computation and various neural network layers.

Note

ndarray provides almost the same routines as symbol. Most routines between these two packages share the source code. But ndarray differs from symbol in few aspects:

  • ndarray adopts imperative programming, namely sentences are executed step-by-step so that the results can be obtained immediately whereas symbol adopts declarative programming.
  • Most binary operators in ndarray such as + and > have broadcasting enabled by default.

In the rest of this document, we first overview the methods provided by the ndarray.NDArray class, and then list other routines provided by the ndarray package.

The ndarray package provides several classes:

NDArray An array object representing a multidimensional, homogeneous array of fixed-size items.
sparse.CSRNDArray A sparse representation of 2D NDArray in the Compressed Sparse Row format.
sparse.RowSparseNDArray A sparse representation of a set of NDArray row slices at given indices.

The NDArray class

Array attributes

NDArray.shape Tuple of array dimensions.
NDArray.size Number of elements in the array.
NDArray.context Device context of the array.
NDArray.dtype Data-type of the array’s elements.
NDArray.stype Storage-type of the array.

Array conversion

NDArray.copy Makes a copy of this NDArray, keeping the same context.
NDArray.copyto Copies the value of this array to another array.
NDArray.as_in_context Returns an array on the target device with the same value as this array.
NDArray.asnumpy Returns a numpy.ndarray object with value copied from this array.
NDArray.asscalar Returns a scalar whose value is copied from this array.
NDArray.astype Returns a copy of the array after casting to a specified type.
NDArray.tostype Return a copy of the array with chosen storage type.

Array creation

NDArray.zeros_like Convenience fluent method for zeros_like().
NDArray.ones_like Convenience fluent method for ones_like().

Array change shape

NDArray.T Returns a copy of the array with axes transposed.
NDArray.shape_array Convenience fluent method for shape_array().
NDArray.size_array Convenience fluent method for size_array().
NDArray.reshape Returns a view of this array with a new shape without altering any data.
NDArray.reshape_like Convenience fluent method for reshape_like().
NDArray.flatten Convenience fluent method for flatten().
NDArray.expand_dims Convenience fluent method for expand_dims().
NDArray.split Convenience fluent method for split().
NDArray.diag Convenience fluent method for diag().

Array expand elements

NDArray.broadcast_to Broadcasts the input array to a new shape.
NDArray.broadcast_axes Convenience fluent method for broadcast_axes().
NDArray.broadcast_like Broadcasts the input array to the shape of other.
NDArray.tile Convenience fluent method for tile().
NDArray.pad Convenience fluent method for pad().

Array rearrange elements

NDArray.transpose Convenience fluent method for transpose().
NDArray.swapaxes Convenience fluent method for swapaxes().
NDArray.flip Convenience fluent method for flip().
NDArray.depth_to_space Convenience fluent method for depth_to_space().
NDArray.space_to_depth Convenience fluent method for space_to_depth().

Array reduction

NDArray.sum Convenience fluent method for sum().
NDArray.nansum Convenience fluent method for nansum().
NDArray.prod Convenience fluent method for prod().
NDArray.nanprod Convenience fluent method for nanprod().
NDArray.mean Convenience fluent method for mean().
NDArray.max Convenience fluent method for max().
NDArray.min Convenience fluent method for min().
NDArray.norm Convenience fluent method for norm().

Array rounding

NDArray.round Convenience fluent method for round().
NDArray.rint Convenience fluent method for rint().
NDArray.fix Convenience fluent method for fix().
NDArray.floor Convenience fluent method for floor().
NDArray.ceil Convenience fluent method for ceil().
NDArray.trunc Convenience fluent method for trunc().

Array sorting and searching

NDArray.sort Convenience fluent method for sort().
NDArray.argsort Convenience fluent method for argsort().
NDArray.topk Convenience fluent method for topk().
NDArray.argmax Convenience fluent method for argmax().
NDArray.argmin Convenience fluent method for argmin().
NDArray.argmax_channel Convenience fluent method for argmax_channel().

Arithmetic operations

NDArray.__add__ x.__add__(y) <=> x+y <=> mx.nd.add(x, y)
NDArray.__sub__ x.__sub__(y) <=> x-y <=> mx.nd.subtract(x, y)
NDArray.__rsub__ x.__rsub__(y) <=> y-x <=> mx.nd.subtract(y, x)
NDArray.__neg__ x.__neg__(y) <=> -x
NDArray.__mul__ x.__mul__(y) <=> x*y <=> mx.nd.multiply(x, y)
NDArray.__div__ x.__div__(y) <=> x/y <=> mx.nd.divide(x, y)
NDArray.__rdiv__ x.__rdiv__(y) <=> y/x <=> mx.nd.divide(y, x)
NDArray.__mod__ x.__mod__(y) <=> x%y <=> mx.nd.modulo(x, y)
NDArray.__rmod__ x.__rmod__(y) <=> y%x <=> mx.nd.modulo(y, x)
NDArray.__pow__ x.__pow__(y) <=> x**y <=> mx.nd.power(x,y)

Trigonometric functions

NDArray.sin Convenience fluent method for sin().
NDArray.cos Convenience fluent method for cos().
NDArray.tan Convenience fluent method for tan().
NDArray.arcsin Convenience fluent method for arcsin().
NDArray.arccos Convenience fluent method for arccos().
NDArray.arctan Convenience fluent method for arctan().
NDArray.degrees Convenience fluent method for degrees().
NDArray.radians Convenience fluent method for radians().

Hyperbolic functions

NDArray.sinh Convenience fluent method for sinh().
NDArray.cosh Convenience fluent method for cosh().
NDArray.tanh Convenience fluent method for tanh().
NDArray.arcsinh Convenience fluent method for arcsinh().
NDArray.arccosh Convenience fluent method for arccosh().
NDArray.arctanh Convenience fluent method for arctanh().

Exponents and logarithms

NDArray.exp Convenience fluent method for exp().
NDArray.expm1 Convenience fluent method for expm1().
NDArray.log Convenience fluent method for log().
NDArray.log10 Convenience fluent method for log10().
NDArray.log2 Convenience fluent method for log2().
NDArray.log1p Convenience fluent method for log1p().

Powers

NDArray.sqrt Convenience fluent method for sqrt().
NDArray.rsqrt Convenience fluent method for rsqrt().
NDArray.cbrt Convenience fluent method for cbrt().
NDArray.rcbrt Convenience fluent method for rcbrt().
NDArray.square Convenience fluent method for square().
NDArray.reciprocal Convenience fluent method for reciprocal().

Basic neural network functions

NDArray.relu Convenience fluent method for relu().
NDArray.sigmoid Convenience fluent method for sigmoid().
NDArray.softmax Convenience fluent method for softmax().
NDArray.log_softmax Convenience fluent method for log_softmax().

In-place arithmetic operations

NDArray.__iadd__ x.__iadd__(y) <=> x+=y
NDArray.__isub__ x.__isub__(y) <=> x-=y
NDArray.__imul__ x.__imul__(y) <=> x*=y
NDArray.__idiv__ x.__rdiv__(y) <=> x/=y
NDArray.__imod__ x.__rmod__(y) <=> x%=y

Comparison operators

NDArray.__lt__ x.__lt__(y) <=> x mx.nd.lesser(x, y)
NDArray.__le__ x.__le__(y) <=> x<=y <=> mx.nd.less_equal(x, y)
NDArray.__gt__ x.__gt__(y) <=> x>y <=> mx.nd.greater(x, y)
NDArray.__ge__ x.__ge__(y) <=> x>=y <=> mx.nd.greater_equal(x, y)
NDArray.__eq__ x.__eq__(y) <=> x==y <=> mx.nd.equal(x, y)
NDArray.__ne__ x.__ne__(y) <=> x!=y <=> mx.nd.not_equal(x, y)

Indexing

NDArray.__getitem__ x.__getitem__(i) <=> x[i]
NDArray.__setitem__ x.__setitem__(i, y) <=> x[i]=y
NDArray.slice Convenience fluent method for slice().
NDArray.slice_axis Convenience fluent method for slice_axis().
NDArray.slice_like Convenience fluent method for slice_like().
NDArray.take Convenience fluent method for take().
NDArray.one_hot Convenience fluent method for one_hot().
NDArray.pick Convenience fluent method for pick().

Lazy evaluation

NDArray.wait_to_read Waits until all previous write operations on the current array are finished.

Miscellaneous

NDArray.clip Convenience fluent method for clip().
NDArray.sign Convenience fluent method for sign().

Array creation routines

array Creates an array from any object exposing the array interface.
empty Returns a new array of given shape and type, without initializing entries.
zeros Return a new array of given shape and type, filled with zeros.
zeros_like Return an array of zeros with the same shape, type and storage type as the input array.
ones Returns a new array filled with all ones, with the given shape and type.
ones_like Return an array of ones with the same shape and type as the input array.
full Returns a new array of given shape and type, filled with the given value val.
arange Returns evenly spaced values within a given interval.
diag Extracts a diagonal or constructs a diagonal array.
load Loads an array from file.
save Saves a list of arrays or a dict of str->array to file.

Array manipulation routines

Changing array shape and type

cast Casts all elements of the input to a new type.
shape_array Returns a 1D int64 array containing the shape of data.
size_array Returns a 1D int64 array containing the size of data.
reshape Reshapes the input array.
reshape_like Reshape some or all dimensions of lhs to have the same shape as some or all dimensions of rhs.
flatten Flattens the input array into a 2-D array by collapsing the higher dimensions.
expand_dims Inserts a new axis of size 1 into the array shape

Expanding array elements

broadcast_to Broadcasts the input array to a new shape.
broadcast_axes Broadcasts the input array over particular axes.
broadcast_like Broadcasts lhs to have the same shape as rhs.
repeat Repeats elements of an array.
tile Repeats the whole array multiple times.
pad Pads an input array with a constant or edge values of the array.

Rearranging elements

transpose Permutes the dimensions of an array.
swapaxes Interchanges two axes of an array.
flip Reverses the order of elements along given axis while preserving array shape.
depth_to_space Rearranges(permutes) data from depth into blocks of spatial data.
space_to_depth Rearranges(permutes) blocks of spatial data into depth.

Joining and splitting arrays

concat Joins input arrays along a given axis.
split Splits an array along a particular axis into multiple sub-arrays.
stack Join a sequence of arrays along a new axis.

Indexing routines

slice Slices a region of the array.
slice_axis Slices along a given axis.
slice_like Slices a region of the array like the shape of another array.
take Takes elements from an input array along the given axis.
batch_take Takes elements from a data batch.
one_hot Returns a one-hot array.
pick Picks elements from an input array according to the input indices along the given axis.
where Return the elements, either from x or y, depending on the condition.
ravel_multi_index Converts a batch of index arrays into an array of flat indices.
unravel_index Converts an array of flat indices into a batch of index arrays.

Mathematical functions

Arithmetic operations

add Returns element-wise sum of the input arrays with broadcasting.
subtract Returns element-wise difference of the input arrays with broadcasting.
negative Numerical negative of the argument, element-wise.
multiply Returns element-wise product of the input arrays with broadcasting.
divide Returns element-wise division of the input arrays with broadcasting.
modulo Returns element-wise modulo of the input arrays with broadcasting.
dot Dot product of two arrays.
batch_dot Batchwise dot product.
add_n Adds all input arguments element-wise.

Trigonometric functions

sin Computes the element-wise sine of the input array.
cos Computes the element-wise cosine of the input array.
tan Computes the element-wise tangent of the input array.
arcsin Returns element-wise inverse sine of the input array.
arccos Returns element-wise inverse cosine of the input array.
arctan Returns element-wise inverse tangent of the input array.
broadcast_hypot Returns the hypotenuse of a right angled triangle, given its “legs” with broadcasting.
degrees Converts each element of the input array from radians to degrees.
radians Converts each element of the input array from degrees to radians.

Hyperbolic functions

sinh Returns the hyperbolic sine of the input array, computed element-wise.
cosh Returns the hyperbolic cosine of the input array, computed element-wise.
tanh Returns the hyperbolic tangent of the input array, computed element-wise.
arcsinh Returns the element-wise inverse hyperbolic sine of the input array, computed element-wise.
arccosh Returns the element-wise inverse hyperbolic cosine of the input array, computed element-wise.
arctanh Returns the element-wise inverse hyperbolic tangent of the input array, computed element-wise.

Reduce functions

sum Computes the sum of array elements over given axes.
nansum Computes the sum of array elements over given axes treating Not a Numbers (NaN) as zero.
prod Computes the product of array elements over given axes.
nanprod Computes the product of array elements over given axes treating Not a Numbers (NaN) as one.
mean Computes the mean of array elements over given axes.
max Computes the max of array elements over given axes.
min Computes the min of array elements over given axes.
norm Computes the norm on an NDArray.

Rounding

round Returns element-wise rounded value to the nearest integer of the input.
rint Returns element-wise rounded value to the nearest integer of the input.
fix Returns element-wise rounded value to the nearest integer towards zero of the input.
floor Returns element-wise floor of the input.
ceil Returns element-wise ceiling of the input.
trunc Return the element-wise truncated value of the input.

Exponents and logarithms

exp Returns element-wise exponential value of the input.
expm1 Returns exp(x) - 1 computed element-wise on the input.
log Returns element-wise Natural logarithmic value of the input.
log10 Returns element-wise Base-10 logarithmic value of the input.
log2 Returns element-wise Base-2 logarithmic value of the input.
log1p Returns element-wise log(1 + x) value of the input.

Powers

power Returns result of first array elements raised to powers from second array, element-wise with broadcasting.
sqrt Returns element-wise square-root value of the input.
rsqrt Returns element-wise inverse square-root value of the input.
cbrt Returns element-wise cube-root value of the input.
rcbrt Returns element-wise inverse cube-root value of the input.
square Returns element-wise squared value of the input.
reciprocal Returns the reciprocal of the argument, element-wise.

Comparison

equal Returns the result of element-wise equal to (==) comparison operation with broadcasting.
not_equal Returns the result of element-wise not equal to (!=) comparison operation with broadcasting.
greater Returns the result of element-wise greater than (>) comparison operation with broadcasting.
greater_equal Returns the result of element-wise greater than or equal to (>=) comparison operation with broadcasting.
lesser Returns the result of element-wise lesser than (<) comparison operation with broadcasting.
lesser_equal Returns the result of element-wise lesser than or equal to (<=) comparison operation with broadcasting.

Logical operators

logical_and Returns the result of element-wise logical and comparison operation with broadcasting.
logical_or Returns the result of element-wise logical or comparison operation with broadcasting.
logical_xor Returns the result of element-wise logical xor comparison operation with broadcasting.
logical_not Returns the result of logical NOT (!) function

Random sampling

random.exponential Draw samples from an exponential distribution.
random.gamma Draw random samples from a gamma distribution.
random.generalized_negative_binomial Draw random samples from a generalized negative binomial distribution.
random.multinomial Concurrent sampling from multiple multinomial distributions.
random.negative_binomial Draw random samples from a negative binomial distribution.
random.normal Draw random samples from a normal (Gaussian) distribution.
random.poisson Draw random samples from a Poisson distribution.
random.randint Draw random samples from a discrete uniform distribution.
random.randn Draw random samples from a normal (Gaussian) distribution.
random.shuffle Shuffle the elements randomly.
random.uniform Draw random samples from a uniform distribution.
mxnet.random.seed Seeds the random number generators in MXNet.

Sorting and searching

sort Returns a sorted copy of an input array along the given axis.
topk Returns the top k elements in an input array along the given axis.
argsort Returns the indices that would sort an input array along the given axis.
argmax Returns indices of the maximum values along an axis.
argmin Returns indices of the minimum values along an axis.

Sequence operation

SequenceLast Takes the last element of a sequence.
SequenceMask Sets all elements outside the sequence to a constant value.
SequenceReverse Reverses the elements of each sequence.

Miscellaneous

maximum Returns element-wise maximum of the input arrays with broadcasting.
minimum Returns element-wise minimum of the input arrays with broadcasting.
clip Clips (limits) the values in an array.
abs Returns element-wise absolute value of the input.
sign Returns element-wise sign of the input.
gamma Returns the gamma function (extension of the factorial function to the reals), computed element-wise on the input array.
gammaln Returns element-wise log of the absolute value of the gamma function of the input.

Neural network

Basic

FullyConnected Applies a linear transformation: \(Y = XW^T + b\).
Convolution Compute N-D convolution on (N+2)-D input.
Activation Applies an activation function element-wise to the input.
BatchNorm Batch normalization.
Pooling Performs pooling on the input.
SoftmaxOutput Computes the gradient of cross entropy loss with respect to softmax output.
softmax Applies the softmax function.
log_softmax Computes the log softmax of the input.
relu Computes rectified linear activation.
sigmoid Computes sigmoid of x element-wise.
erf Returns element-wise gauss error function of the input.
erfinv Returns element-wise inverse gauss error function of the input.

More

Correlation Applies correlation to inputs.
Deconvolution Computes 1D or 2D transposed convolution (aka fractionally strided convolution) of the input tensor.
RNN Applies recurrent layers to input data.
Embedding Maps integer indices to vector representations (embeddings).
LeakyReLU Applies Leaky rectified linear unit activation element-wise to the input.
InstanceNorm Applies instance normalization to the n-dimensional input array.
LayerNorm Layer normalization.
L2Normalization Normalize the input array using the L2 norm.
LRN Applies local response normalization to the input.
ROIPooling Performs region of interest(ROI) pooling on the input array.
SoftmaxActivation Applies softmax activation to input.
Dropout Applies dropout operation to input array.
BilinearSampler Applies bilinear sampling to input feature map.
GridGenerator Generates 2D sampling grid for bilinear sampling.
UpSampling Upsamples the given input data.
SpatialTransformer Applies a spatial transformer to input feature map.
LinearRegressionOutput Computes and optimizes for squared loss during backward propagation.
LogisticRegressionOutput Applies a logistic function to the input.
MAERegressionOutput Computes mean absolute error of the input.
SVMOutput Computes support vector machine based transformation of the input.
softmax_cross_entropy Calculate cross entropy of softmax output and one-hot label.
smooth_l1 Calculate Smooth L1 Loss(lhs, scalar) by summing
IdentityAttachKLSparseReg Apply a sparse regularization to the output a sigmoid activation function.
MakeLoss Make your own loss function in network construction.
BlockGrad Stops gradient computation.
Custom Apply a custom operator implemented in a frontend language (like Python).

API Reference

class mxnet.ndarray.NDArray(handle, writable=True)[source]

An array object representing a multidimensional, homogeneous array of fixed-size items.

__repr__()[source]

Returns a string representation of the array.

__add__(other)[source]

x.__add__(y) <=> x+y <=> mx.nd.add(x, y)

__iadd__(other)[source]

x.__iadd__(y) <=> x+=y

__sub__(other)[source]

x.__sub__(y) <=> x-y <=> mx.nd.subtract(x, y)

__isub__(other)[source]

x.__isub__(y) <=> x-=y

__rsub__(other)[source]

x.__rsub__(y) <=> y-x <=> mx.nd.subtract(y, x)

__mul__(other)[source]

x.__mul__(y) <=> x*y <=> mx.nd.multiply(x, y)

__neg__()[source]

x.__neg__(y) <=> -x

__imul__(other)[source]

x.__imul__(y) <=> x*=y

__div__(other)[source]

x.__div__(y) <=> x/y <=> mx.nd.divide(x, y)

__rdiv__(other)[source]

x.__rdiv__(y) <=> y/x <=> mx.nd.divide(y, x)

__idiv__(other)[source]

x.__rdiv__(y) <=> x/=y

__mod__(other)[source]

x.__mod__(y) <=> x%y <=> mx.nd.modulo(x, y)

__rmod__(other)[source]

x.__rmod__(y) <=> y%x <=> mx.nd.modulo(y, x)

__imod__(other)[source]

x.__rmod__(y) <=> x%=y

__pow__(other)[source]

x.__pow__(y) <=> x**y <=> mx.nd.power(x,y)

__rpow__(other)[source]

x.__pow__(y) <=> y**x <=> mx.nd.power(y,x)

__eq__(other)[source]

x.__eq__(y) <=> x==y <=> mx.nd.equal(x, y)

__hash__()[source]

Default hash function.

__ne__(other)[source]

x.__ne__(y) <=> x!=y <=> mx.nd.not_equal(x, y)

__gt__(other)[source]

x.__gt__(y) <=> x>y <=> mx.nd.greater(x, y)

__ge__(other)[source]

x.__ge__(y) <=> x>=y <=> mx.nd.greater_equal(x, y)

__lt__(other)[source]

x.__lt__(y) <=> x mx.nd.lesser(x, y)

__le__(other)[source]

x.__le__(y) <=> x<=y <=> mx.nd.less_equal(x, y)

__len__()[source]

Number of element along the first axis.

__setitem__(key, value)[source]

x.__setitem__(i, y) <=> x[i]=y

Sets value to self[key]. This functions supports advanced indexing defined in the following reference with some restrictions.

https://docs.scipy.org/doc/numpy-1.13.0/reference/arrays.indexing.html#combining-advanced-and-basic-indexing

  • If key is a list type, only a list of integers is supported, e.g. key=[1, 2] is supported, while not for key=[[1, 2]].
  • Ellipsis (...) and np.newaxis are not supported.
  • Boolean array indexing is not supported.
Parameters:
  • key (int, mxnet.ndarray.slice, list, np.ndarray, NDArray, or tuple of all previous types) – The indexing key.
  • value (scalar or array-like object that can be broadcast to the shape of self[key]) – The value to set.

Examples

>>> x = mx.nd.zeros((2,3))
>>> x[:] = 1
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> x[:,1:2] = 2
>>> x.asnumpy()
array([[ 1.,  2.,  1.],
       [ 1.,  2.,  1.]], dtype=float32)
>>> x[1:2,1:] = 3
>>> x.asnumpy()
array([[ 1.,  2.,  1.],
       [ 1.,  3.,  3.]], dtype=float32)
>>> x[1:,0:2] = mx.nd.zeros((1,2))
>>> x.asnumpy()
array([[ 1.,  2.,  1.],
       [ 0.,  0.,  3.]], dtype=float32)
>>> x[1,2] = 4
>>> x.asnumpy()
array([[ 1.,  2.,  1.],
       [ 0.,  0.,  4.]], dtype=float32)
>>> x[[0], [1, 2]] = 5
>>> x.asnumpy()
array([[ 1.,  5.,  5.],
       [ 0.,  0.,  4.]], dtype=float32)
>>> x[::-1, 0:2:2] = [6]
>>> x.asnumpy()
array([[ 6.,  5.,  5.],
       [ 6.,  0.,  4.]], dtype=float32)
__getitem__(key)[source]

x.__getitem__(i) <=> x[i]

Returns a sliced view of this array if the elements fetched are contiguous in memory; otherwise, returns a newly created NDArray. This functions supports advanced indexing defined in the following reference with some restrictions.

https://docs.scipy.org/doc/numpy-1.13.0/reference/arrays.indexing.html#combining-advanced-and-basic-indexing

  • If key is a list type, only a list of integers is supported, e.g. key=[1, 2] is supported, while not for key=[[1, 2]].
  • Ellipsis (...) and np.newaxis are not supported.
  • Boolean array indexing is not supported.
Parameters:key (int, mxnet.ndarray.slice, list, np.ndarray, NDArray, or tuple of all previous types) – Indexing key.

Examples

>>> x = mx.nd.arange(0,6).reshape((2,3))
>>> x.asnumpy()
array([[ 0.,  1.,  2.],
       [ 3.,  4.,  5.]], dtype=float32)
>>> x[1].asnumpy()
array([ 3.,  4.,  5.], dtype=float32)
>>> y = x[0:1]
>>> y[:] = 2
>>> x.asnumpy()
array([[ 2.,  2.,  2.],
       [ 3.,  4.,  5.]], dtype=float32)
>>> x = mx.nd.arange(0, 8, dtype='int32').reshape((2, 2, 2))
>>> x[[0, 1]]
[[[0 1]
  [2 3]]
 [[4 5]
  [6 7]]]
>>> x[1:, [0, 1]]
[[[4 5]
  [6 7]]]
>>> y = np.array([0, 1], dtype='int32')
>>> x[1:, y]
[[[4 5]
  [6 7]]]
>>> y = mx.nd.array([0, 1], dtype='int32')
>>> x[1:, y]
[[[4 5]
  [6 7]]]
reshape(*shape, **kwargs)[source]

Returns a view of this array with a new shape without altering any data.

Parameters:
  • shape (tuple of int, or n ints) –

    The new shape should not change the array size, namely np.prod(new_shape) should be equal to np.prod(self.shape). Some dimensions of the shape can take special values from the set {0, -1, -2, -3, -4}. The significance of each is explained below:

    • 0 copy this dimension from the input to the output shape.

      Example:

      - input shape = (2,3,4), shape = (4,0,2), output shape = (4,3,2)
      - input shape = (2,3,4), shape = (2,0,0), output shape = (2,3,4)
      
    • -1 infers the dimension of the output shape by using the remainder of the input dimensions keeping the size of the new array same as that of the input array. At most one dimension of shape can be -1.

      Example:

      - input shape = (2,3,4), shape = (6,1,-1), output shape = (6,1,4)
      - input shape = (2,3,4), shape = (3,-1,8), output shape = (3,1,8)
      - input shape = (2,3,4), shape=(-1,), output shape = (24,)
      
    • -2 copy all/remainder of the input dimensions to the output shape.

      Example:

      - input shape = (2,3,4), shape = (-2,), output shape = (2,3,4)
      - input shape = (2,3,4), shape = (2,-2), output shape = (2,3,4)
      - input shape = (2,3,4), shape = (-2,1,1), output shape = (2,3,4,1,1)
      
    • -3 use the product of two consecutive dimensions of the input shape as the output dimension.

      Example:

      - input shape = (2,3,4), shape = (-3,4), output shape = (6,4)
      - input shape = (2,3,4,5), shape = (-3,-3), output shape = (6,20)
      - input shape = (2,3,4), shape = (0,-3), output shape = (2,12)
      - input shape = (2,3,4), shape = (-3,-2), output shape = (6,4)
      
    • -4 split one dimension of the input into two dimensions passed subsequent to -4 in shape (can contain -1).

      Example:

      - input shape = (2,3,4), shape = (-4,1,2,-2), output shape =(1,2,3,4)
      - input shape = (2,3,4), shape = (2,-4,-1,3,-2), output shape = (2,1,3,4)
      
    • If the argument reverse is set to 1, then the special values are inferred from right to left.

      Example:

      - without reverse=1, for input shape = (10,5,4), shape = (-1,0), output shape would be                 (40,5).
      - with reverse=1, output shape will be (50,4).
      
  • reverse (bool, default False) – If true then the special values are inferred from right to left. Only supported as keyword argument.
Returns:

An array with desired shape that shares data with this array.

Return type:

NDArray

Examples

>>> x = mx.nd.arange(0,6).reshape(2,3)
>>> x.asnumpy()
array([[ 0.,  1.,  2.],
       [ 3.,  4.,  5.]], dtype=float32)
>>> y = x.reshape(3,2)
>>> y.asnumpy()
array([[ 0.,  1.],
       [ 2.,  3.],
       [ 4.,  5.]], dtype=float32)
>>> y = x.reshape(3,-1)
>>> y.asnumpy()
array([[ 0.,  1.],
       [ 2.,  3.],
       [ 4.,  5.]], dtype=float32)
>>> y = x.reshape(3,2)
>>> y.asnumpy()
array([[ 0.,  1.],
       [ 2.,  3.],
       [ 4.,  5.]], dtype=float32)
>>> y = x.reshape(-3)
>>> y.asnumpy()
array([ 0.  1.  2.  3.  4.  5.], dtype=float32)
>>> y[:] = -1
>>> x.asnumpy()
array([[-1., -1., -1.],
       [-1., -1., -1.]], dtype=float32)
reshape_like(*args, **kwargs)[source]

Convenience fluent method for reshape_like().

The arguments are the same as for reshape_like(), with this array as data.

zeros_like(*args, **kwargs)[source]

Convenience fluent method for zeros_like().

The arguments are the same as for zeros_like(), with this array as data.

ones_like(*args, **kwargs)[source]

Convenience fluent method for ones_like().

The arguments are the same as for ones_like(), with this array as data.

broadcast_axes(*args, **kwargs)[source]

Convenience fluent method for broadcast_axes().

The arguments are the same as for broadcast_axes(), with this array as data.

repeat(*args, **kwargs)[source]

Convenience fluent method for repeat().

The arguments are the same as for repeat(), with this array as data.

pad(*args, **kwargs)[source]

Convenience fluent method for pad().

The arguments are the same as for pad(), with this array as data.

swapaxes(*args, **kwargs)[source]

Convenience fluent method for swapaxes().

The arguments are the same as for swapaxes(), with this array as data.

split(*args, **kwargs)[source]

Convenience fluent method for split().

The arguments are the same as for split(), with this array as data.

split_v2(*args, **kwargs)[source]

Convenience fluent method for split_v2().

The arguments are the same as for split_v2(), with this array as data.

slice(*args, **kwargs)[source]

Convenience fluent method for slice().

The arguments are the same as for slice(), with this array as data.

slice_axis(*args, **kwargs)[source]

Convenience fluent method for slice_axis().

The arguments are the same as for slice_axis(), with this array as data.

slice_like(*args, **kwargs)[source]

Convenience fluent method for slice_like().

The arguments are the same as for slice_like(), with this array as data.

take(*args, **kwargs)[source]

Convenience fluent method for take().

The arguments are the same as for take(), with this array as data.

one_hot(*args, **kwargs)[source]

Convenience fluent method for one_hot().

The arguments are the same as for one_hot(), with this array as data.

pick(*args, **kwargs)[source]

Convenience fluent method for pick().

The arguments are the same as for pick(), with this array as data.

sort(*args, **kwargs)[source]

Convenience fluent method for sort().

The arguments are the same as for sort(), with this array as data.

topk(*args, **kwargs)[source]

Convenience fluent method for topk().

The arguments are the same as for topk(), with this array as data.

argsort(*args, **kwargs)[source]

Convenience fluent method for argsort().

The arguments are the same as for argsort(), with this array as data.

argmax(*args, **kwargs)[source]

Convenience fluent method for argmax().

The arguments are the same as for argmax(), with this array as data.

argmax_channel(*args, **kwargs)[source]

Convenience fluent method for argmax_channel().

The arguments are the same as for argmax_channel(), with this array as data.

argmin(*args, **kwargs)[source]

Convenience fluent method for argmin().

The arguments are the same as for argmin(), with this array as data.

clip(*args, **kwargs)[source]

Convenience fluent method for clip().

The arguments are the same as for clip(), with this array as data.

abs(*args, **kwargs)[source]

Convenience fluent method for abs().

The arguments are the same as for abs(), with this array as data.

sign(*args, **kwargs)[source]

Convenience fluent method for sign().

The arguments are the same as for sign(), with this array as data.

flatten(*args, **kwargs)[source]

Convenience fluent method for flatten().

The arguments are the same as for flatten(), with this array as data.

shape_array(*args, **kwargs)[source]

Convenience fluent method for shape_array().

The arguments are the same as for shape_array(), with this array as data.

size_array(*args, **kwargs)[source]

Convenience fluent method for size_array().

The arguments are the same as for size_array(), with this array as data.

expand_dims(*args, **kwargs)[source]

Convenience fluent method for expand_dims().

The arguments are the same as for expand_dims(), with this array as data.

tile(*args, **kwargs)[source]

Convenience fluent method for tile().

The arguments are the same as for tile(), with this array as data.

transpose(*args, **kwargs)[source]

Convenience fluent method for transpose().

The arguments are the same as for transpose(), with this array as data.

flip(*args, **kwargs)[source]

Convenience fluent method for flip().

The arguments are the same as for flip(), with this array as data.

depth_to_space(*args, **kwargs)[source]

Convenience fluent method for depth_to_space().

The arguments are the same as for depth_to_space(), with this array as data.

space_to_depth(*args, **kwargs)[source]

Convenience fluent method for space_to_depth().

The arguments are the same as for space_to_depth(), with this array as data.

diag(k=0, **kwargs)[source]

Convenience fluent method for diag().

The arguments are the same as for diag(), with this array as data.

sum(*args, **kwargs)[source]

Convenience fluent method for sum().

The arguments are the same as for sum(), with this array as data.

nansum(*args, **kwargs)[source]

Convenience fluent method for nansum().

The arguments are the same as for nansum(), with this array as data.

prod(*args, **kwargs)[source]

Convenience fluent method for prod().

The arguments are the same as for prod(), with this array as data.

nanprod(*args, **kwargs)[source]

Convenience fluent method for nanprod().

The arguments are the same as for nanprod(), with this array as data.

mean(*args, **kwargs)[source]

Convenience fluent method for mean().

The arguments are the same as for mean(), with this array as data.

max(*args, **kwargs)[source]

Convenience fluent method for max().

The arguments are the same as for max(), with this array as data.

min(*args, **kwargs)[source]

Convenience fluent method for min().

The arguments are the same as for min(), with this array as data.

norm(*args, **kwargs)[source]

Convenience fluent method for norm().

The arguments are the same as for norm(), with this array as data.

round(*args, **kwargs)[source]

Convenience fluent method for round().

The arguments are the same as for round(), with this array as data.

rint(*args, **kwargs)[source]

Convenience fluent method for rint().

The arguments are the same as for rint(), with this array as data.

fix(*args, **kwargs)[source]

Convenience fluent method for fix().

The arguments are the same as for fix(), with this array as data.

floor(*args, **kwargs)[source]

Convenience fluent method for floor().

The arguments are the same as for floor(), with this array as data.

ceil(*args, **kwargs)[source]

Convenience fluent method for ceil().

The arguments are the same as for ceil(), with this array as data.

trunc(*args, **kwargs)[source]

Convenience fluent method for trunc().

The arguments are the same as for trunc(), with this array as data.

sin(*args, **kwargs)[source]

Convenience fluent method for sin().

The arguments are the same as for sin(), with this array as data.

cos(*args, **kwargs)[source]

Convenience fluent method for cos().

The arguments are the same as for cos(), with this array as data.

tan(*args, **kwargs)[source]

Convenience fluent method for tan().

The arguments are the same as for tan(), with this array as data.

arcsin(*args, **kwargs)[source]

Convenience fluent method for arcsin().

The arguments are the same as for arcsin(), with this array as data.

arccos(*args, **kwargs)[source]

Convenience fluent method for arccos().

The arguments are the same as for arccos(), with this array as data.

arctan(*args, **kwargs)[source]

Convenience fluent method for arctan().

The arguments are the same as for arctan(), with this array as data.

degrees(*args, **kwargs)[source]

Convenience fluent method for degrees().

The arguments are the same as for degrees(), with this array as data.

radians(*args, **kwargs)[source]

Convenience fluent method for radians().

The arguments are the same as for radians(), with this array as data.

sinh(*args, **kwargs)[source]

Convenience fluent method for sinh().

The arguments are the same as for sinh(), with this array as data.

cosh(*args, **kwargs)[source]

Convenience fluent method for cosh().

The arguments are the same as for cosh(), with this array as data.

tanh(*args, **kwargs)[source]

Convenience fluent method for tanh().

The arguments are the same as for tanh(), with this array as data.

arcsinh(*args, **kwargs)[source]

Convenience fluent method for arcsinh().

The arguments are the same as for arcsinh(), with this array as data.

arccosh(*args, **kwargs)[source]

Convenience fluent method for arccosh().

The arguments are the same as for arccosh(), with this array as data.

arctanh(*args, **kwargs)[source]

Convenience fluent method for arctanh().

The arguments are the same as for arctanh(), with this array as data.

exp(*args, **kwargs)[source]

Convenience fluent method for exp().

The arguments are the same as for exp(), with this array as data.

expm1(*args, **kwargs)[source]

Convenience fluent method for expm1().

The arguments are the same as for expm1(), with this array as data.

log(*args, **kwargs)[source]

Convenience fluent method for log().

The arguments are the same as for log(), with this array as data.

log10(*args, **kwargs)[source]

Convenience fluent method for log10().

The arguments are the same as for log10(), with this array as data.

log2(*args, **kwargs)[source]

Convenience fluent method for log2().

The arguments are the same as for log2(), with this array as data.

log1p(*args, **kwargs)[source]

Convenience fluent method for log1p().

The arguments are the same as for log1p(), with this array as data.

sqrt(*args, **kwargs)[source]

Convenience fluent method for sqrt().

The arguments are the same as for sqrt(), with this array as data.

rsqrt(*args, **kwargs)[source]

Convenience fluent method for rsqrt().

The arguments are the same as for rsqrt(), with this array as data.

cbrt(*args, **kwargs)[source]

Convenience fluent method for cbrt().

The arguments are the same as for cbrt(), with this array as data.

rcbrt(*args, **kwargs)[source]

Convenience fluent method for rcbrt().

The arguments are the same as for rcbrt(), with this array as data.

square(*args, **kwargs)[source]

Convenience fluent method for square().

The arguments are the same as for square(), with this array as data.

reciprocal(*args, **kwargs)[source]

Convenience fluent method for reciprocal().

The arguments are the same as for reciprocal(), with this array as data.

relu(*args, **kwargs)[source]

Convenience fluent method for relu().

The arguments are the same as for relu(), with this array as data.

sigmoid(*args, **kwargs)[source]

Convenience fluent method for sigmoid().

The arguments are the same as for sigmoid(), with this array as data.

softmax(*args, **kwargs)[source]

Convenience fluent method for softmax().

The arguments are the same as for softmax(), with this array as data.

log_softmax(*args, **kwargs)[source]

Convenience fluent method for log_softmax().

The arguments are the same as for log_softmax(), with this array as data.

softmin(*args, **kwargs)[source]

Convenience fluent method for softmin().

The arguments are the same as for softmin(), with this array as data.

squeeze(*args, **kwargs)[source]

Convenience fluent method for squeeze().

The arguments are the same as for squeeze(), with this array as data.

broadcast_to(shape)[source]

Broadcasts the input array to a new shape.

Broadcasting is only allowed on axes with size 1. The new shape cannot change the number of dimensions. For example, you could broadcast from shape (2, 1) to (2, 3), but not from shape (2, 3) to (2, 3, 3).

Parameters:shape (tuple of int) – The shape of the desired array.
Returns:A NDArray with the desired shape that is not sharing data with this array, even if the new shape is the same as self.shape.
Return type:NDArray

Examples

>>> x = mx.nd.arange(0,3).reshape((1,3,1))
>>> x.asnumpy()
array([[[ 0.],
        [ 1.],
        [ 2.]]], dtype=float32)
>>> y = x.broadcast_to((2,3,3))
>>> y.asnumpy()
array([[[ 0.,  0.,  0.],
        [ 1.,  1.,  1.],
        [ 2.,  2.,  2.]],

       [[ 0.,  0.,  0.],
        [ 1.,  1.,  1.],
        [ 2.,  2.,  2.]]], dtype=float32)
broadcast_like(other)[source]

Broadcasts the input array to the shape of other.

Broadcasting is only allowed on axes with size 1. The new shape cannot change the number of dimensions. For example, you could broadcast from shape (2, 1) to (2, 3), but not from shape (2, 3) to (2, 3, 3).

Parameters:other (NDArray) – Array with shape of the desired array.
Returns:A NDArray with the desired shape that is not sharing data with this array, even if the new shape is the same as self.shape.
Return type:NDArray

Examples

>>> x = mx.nd.arange(0,3).reshape((1,3,1))
>>> x.asnumpy()
array([[[ 0.],
        [ 1.],
        [ 2.]]], dtype=float32)
>>> y = x.broadcast_like(mx.nd.ones((2,3,3)))
>>> y.asnumpy()
array([[[ 0.,  0.,  0.],
        [ 1.,  1.,  1.],
        [ 2.,  2.,  2.]],

       [[ 0.,  0.,  0.],
        [ 1.,  1.,  1.],
        [ 2.,  2.,  2.]]], dtype=float32)
wait_to_read()[source]

Waits until all previous write operations on the current array are finished.

This method guarantees that all previous write operations that pushed into the backend engine for execution are actually finished.

Examples

>>> import time
>>> tic = time.time()
>>> a = mx.nd.ones((1000,1000))
>>> b = mx.nd.dot(a, a)
>>> print(time.time() - tic) 
0.003854036331176758
>>> b.wait_to_read()
>>> print(time.time() - tic) 
0.0893700122833252
ndim

Returns the number of dimensions of this array

Examples

>>> x = mx.nd.array([1, 2, 3, 4])
>>> x.ndim
1
>>> x = mx.nd.array([[1, 2], [3, 4]])
>>> x.ndim
2
shape

Tuple of array dimensions.

Examples

>>> x = mx.nd.array([1, 2, 3, 4])
>>> x.shape
(4L,)
>>> y = mx.nd.zeros((2, 3, 4))
>>> y.shape
(2L, 3L, 4L)
size

Number of elements in the array.

Equivalent to the product of the array’s dimensions.

Examples

>>> import numpy as np
>>> x = mx.nd.zeros((3, 5, 2))
>>> x.size
30
>>> np.prod(x.shape)
30
context

Device context of the array.

Examples

>>> x = mx.nd.array([1, 2, 3, 4])
>>> x.context
cpu(0)
>>> type(x.context)

>>> y = mx.nd.zeros((2,3), mx.gpu(0))
>>> y.context
gpu(0)
dtype

Data-type of the array’s elements.

Returns:This NDArray’s data type.
Return type:numpy.dtype

Examples

>>> x = mx.nd.zeros((2,3))
>>> x.dtype

>>> y = mx.nd.zeros((2,3), dtype='int32')
>>> y.dtype

stype

Storage-type of the array.

T

Returns a copy of the array with axes transposed.

Equivalent to mx.nd.transpose(self) except that self is returned if self.ndim < 2.

Unlike numpy.ndarray.T, this function returns a copy rather than a view of the array unless self.ndim < 2.

Examples

>>> x = mx.nd.arange(0,6).reshape((2,3))
>>> x.asnumpy()
array([[ 0.,  1.,  2.],
       [ 3.,  4.,  5.]], dtype=float32)
>>> x.T.asnumpy()
array([[ 0.,  3.],
       [ 1.,  4.],
       [ 2.,  5.]], dtype=float32)
asnumpy()[source]

Returns a numpy.ndarray object with value copied from this array.

Examples

>>> x = mx.nd.ones((2,3))
>>> y = x.asnumpy()
>>> type(y)

>>> y
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> z = mx.nd.ones((2,3), dtype='int32')
>>> z.asnumpy()
array([[1, 1, 1],
       [1, 1, 1]], dtype=int32)
asscalar()[source]

Returns a scalar whose value is copied from this array.

This function is equivalent to self.asnumpy()[0]. This NDArray must have shape (1,).

Examples

>>> x = mx.nd.ones((1,), dtype='int32')
>>> x.asscalar()
1
>>> type(x.asscalar())

astype(dtype, copy=True)[source]

Returns a copy of the array after casting to a specified type.

Parameters:
  • dtype (numpy.dtype or str) – The type of the returned array.
  • copy (bool) – Default True. By default, astype always returns a newly allocated ndarray on the same context. If this is set to False, and the dtype requested is the same as the ndarray’s dtype, the ndarray is returned instead of a copy.
Returns:

The copied array after casting to the specified type, or the same array if copy=False and dtype is the same as the input array.

Return type:

NDArray, CSRNDArray or RowSparseNDArray

Examples

>>> x = mx.nd.zeros((2,3), dtype='float32')
>>> y = x.astype('int32')
>>> y.dtype

copyto(other)[source]

Copies the value of this array to another array.

If other is a NDArray object, then other.shape and self.shape should be the same. This function copies the value from self to other.

If other is a context, a new NDArray will be first created on the target context, and the value of self is copied.

Parameters:other (NDArray or Context) – The destination array or context.
Returns:The copied array. If other is an NDArray, then the return value and other will point to the same NDArray.
Return type:NDArray, CSRNDArray or RowSparseNDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.zeros((2,3), mx.gpu(0))
>>> z = x.copyto(y)
>>> z is y
True
>>> y.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.copyto(mx.gpu(0))

copy()[source]

Makes a copy of this NDArray, keeping the same context.

Returns:The copied array
Return type:NDArray, CSRNDArray or RowSparseNDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = x.copy()
>>> y.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
as_in_context(context)[source]

Returns an array on the target device with the same value as this array.

If the target context is the same as self.context, then self is returned. Otherwise, a copy is made.

Parameters:context (Context) – The target context.
Returns:The target array.
Return type:NDArray, CSRNDArray or RowSparseNDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = x.as_in_context(mx.cpu())
>>> y is x
True
>>> z = x.as_in_context(mx.gpu(0))
>>> z is x
False
attach_grad(grad_req='write', stype=None)[source]

Attach a gradient buffer to this NDArray, so that backward can compute gradient with respect to it.

Parameters:
  • grad_req ({'write', 'add', 'null'}) – How gradient will be accumulated. - ‘write’: gradient will be overwritten on every backward. - ‘add’: gradient will be added to existing value on every backward. - ‘null’: do not compute gradient for this NDArray.
  • stype (str, optional) – The storage type of the gradient array. Defaults to the same stype of this NDArray.
grad

Returns gradient buffer attached to this NDArray.

detach()[source]

Returns a new NDArray, detached from the current graph.

backward(out_grad=None, retain_graph=False, train_mode=True)[source]

Compute the gradients of this NDArray w.r.t variables.

Parameters:
  • out_grad (NDArray, optional) – Gradient with respect to head.
  • retain_graph (bool, optional) – Whether to retain the computaion graph for another backward pass on the same graph. By default the computaion history is cleared.
  • train_mode (bool, optional) – Whether to compute gradient for training or inference.
tostype(stype)[source]

Return a copy of the array with chosen storage type.

Returns:A copy of the array with the chosen storage stype
Return type:NDArray, CSRNDArray or RowSparseNDArray
to_dlpack_for_read()[source]

Returns a reference view of NDArray that represents as DLManagedTensor until all previous write operations on the current array are finished.

Returns:a reference view of NDArray that represents as DLManagedTensor.
Return type:PyCapsule (the pointer of DLManagedTensor)

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.to_dlpack_for_read(x)
>>> type(y)

>>> z = mx.nd.from_dlpack(y)
>>> z
[[1. 1. 1.]
 [1. 1. 1.]]

to_dlpack_for_write()[source]

Returns a reference view of NDArray that represents as DLManagedTensor until all previous read/write operations on the current array are finished.

Returns:a reference view of NDArray that represents as DLManagedTensor.
Return type:PyCapsule (the pointer of DLManagedTensor)

Examples

>>> x = mx.nd.ones((2,3))
>>> w = mx.nd.to_dlpack_for_write(x)
>>> type(w)

>>> u = mx.nd.from_dlpack(w)
>>> u += 1
>>> x
[[2. 2. 2.]
 [2. 2. 2.]]

NDArray API of MXNet.

mxnet.ndarray.Activation(data=None, act_type=_Null, out=None, name=None, **kwargs)

Applies an activation function element-wise to the input.

The following activation functions are supported:

  • relu: Rectified Linear Unit, \(y = max(x, 0)\)
  • sigmoid: \(y = \frac{1}{1 + exp(-x)}\)
  • tanh: Hyperbolic tangent, \(y = \frac{exp(x) - exp(-x)}{exp(x) + exp(-x)}\)
  • softrelu: Soft ReLU, or SoftPlus, \(y = log(1 + exp(x))\)
  • softsign: \(y = \frac{x}{1 + abs(x)}\)

Defined in src/operator/nn/activation.cc:L167

Parameters:
  • data (NDArray) – The input array.
  • act_type ({'relu', 'sigmoid', 'softrelu', 'softsign', 'tanh'}, required) – Activation function to be applied.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.BatchNorm(data=None, gamma=None, beta=None, moving_mean=None, moving_var=None, eps=_Null, momentum=_Null, fix_gamma=_Null, use_global_stats=_Null, output_mean_var=_Null, axis=_Null, cudnn_off=_Null, out=None, name=None, **kwargs)

Batch normalization.

Normalizes a data batch by mean and variance, and applies a scale gamma as well as offset beta.

Assume the input has more than one dimension and we normalize along axis 1. We first compute the mean and variance along this axis:

\[\begin{split}data\_mean[i] = mean(data[:,i,:,...]) \\ data\_var[i] = var(data[:,i,:,...])\end{split}\]

Then compute the normalized output, which has the same shape as input, as following:

\[out[:,i,:,...] = \frac{data[:,i,:,...] - data\_mean[i]}{\sqrt{data\_var[i]+\epsilon}} * gamma[i] + beta[i]\]

Both mean and var returns a scalar by treating the input as a vector.

Assume the input has size k on axis 1, then both gamma and beta have shape (k,). If output_mean_var is set to be true, then outputs both data_mean and the inverse of data_var, which are needed for the backward pass. Note that gradient of these two outputs are blocked.

Besides the inputs and the outputs, this operator accepts two auxiliary states, moving_mean and moving_var, which are k-length vectors. They are global statistics for the whole dataset, which are updated by:

moving_mean = moving_mean * momentum + data_mean * (1 - momentum)
moving_var = moving_var * momentum + data_var * (1 - momentum)

If use_global_stats is set to be true, then moving_mean and moving_var are used instead of data_mean and data_var to compute the output. It is often used during inference.

The parameter axis specifies which axis of the input shape denotes the ‘channel’ (separately normalized groups). The default is 1. Specifying -1 sets the channel axis to be the last item in the input shape.

Both gamma and beta are learnable parameters. But if fix_gamma is true, then set gamma to 1 and its gradient to 0.

Note

When fix_gamma is set to True, no sparse support is provided. If fix_gamma is set to False, the sparse tensors will fallback.

Defined in src/operator/nn/batch_norm.cc:L572

Parameters:
  • data (NDArray) – Input data to batch normalization
  • gamma (NDArray) – gamma array
  • beta (NDArray) – beta array
  • moving_mean (NDArray) – running mean of input
  • moving_var (NDArray) – running variance of input
  • eps (double, optional, default=0.0010000000474974513) – Epsilon to prevent div 0. Must be no less than CUDNN_BN_MIN_EPSILON defined in cudnn.h when using cudnn (usually 1e-5)
  • momentum (float, optional, default=0.899999976) – Momentum for moving average
  • fix_gamma (boolean, optional, default=1) – Fix gamma while training
  • use_global_stats (boolean, optional, default=0) – Whether use global moving statistics instead of local batch-norm. This will force change batch-norm into a scale shift operator.
  • output_mean_var (boolean, optional, default=0) – Output the mean and inverse std
  • axis (int, optional, default='1') – Specify which shape axis the channel is specified
  • cudnn_off (boolean, optional, default=0) – Do not select CUDNN operator, if available
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.BatchNorm_v1(data=None, gamma=None, beta=None, eps=_Null, momentum=_Null, fix_gamma=_Null, use_global_stats=_Null, output_mean_var=_Null, out=None, name=None, **kwargs)

Batch normalization.

This operator is DEPRECATED. Perform BatchNorm on the input.

Normalizes a data batch by mean and variance, and applies a scale gamma as well as offset beta.

Assume the input has more than one dimension and we normalize along axis 1. We first compute the mean and variance along this axis:

\[\begin{split}data\_mean[i] = mean(data[:,i,:,...]) \\ data\_var[i] = var(data[:,i,:,...])\end{split}\]

Then compute the normalized output, which has the same shape as input, as following:

\[out[:,i,:,...] = \frac{data[:,i,:,...] - data\_mean[i]}{\sqrt{data\_var[i]+\epsilon}} * gamma[i] + beta[i]\]

Both mean and var returns a scalar by treating the input as a vector.

Assume the input has size k on axis 1, then both gamma and beta have shape (k,). If output_mean_var is set to be true, then outputs both data_mean and data_var as well, which are needed for the backward pass.

Besides the inputs and the outputs, this operator accepts two auxiliary states, moving_mean and moving_var, which are k-length vectors. They are global statistics for the whole dataset, which are updated by:

moving_mean = moving_mean * momentum + data_mean * (1 - momentum)
moving_var = moving_var * momentum + data_var * (1 - momentum)

If use_global_stats is set to be true, then moving_mean and moving_var are used instead of data_mean and data_var to compute the output. It is often used during inference.

Both gamma and beta are learnable parameters. But if fix_gamma is true, then set gamma to 1 and its gradient to 0.

There’s no sparse support for this operator, and it will exhibit problematic behavior if used with sparse tensors.

Defined in src/operator/batch_norm_v1.cc:L95

Parameters:
  • data (NDArray) – Input data to batch normalization
  • gamma (NDArray) – gamma array
  • beta (NDArray) – beta array
  • eps (float, optional, default=0.00100000005) – Epsilon to prevent div 0
  • momentum (float, optional, default=0.899999976) – Momentum for moving average
  • fix_gamma (boolean, optional, default=1) – Fix gamma while training
  • use_global_stats (boolean, optional, default=0) – Whether use global moving statistics instead of local batch-norm. This will force change batch-norm into a scale shift operator.
  • output_mean_var (boolean, optional, default=0) – Output All,normal mean and var
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.BilinearSampler(data=None, grid=None, cudnn_off=_Null, out=None, name=None, **kwargs)

Applies bilinear sampling to input feature map.

Bilinear Sampling is the key of [NIPS2015] “Spatial Transformer Networks”. The usage of the operator is very similar to remap function in OpenCV, except that the operator has the backward pass.

Given \(data\) and \(grid\), then the output is computed by

\[\begin{split}x_{src} = grid[batch, 0, y_{dst}, x_{dst}] \\ y_{src} = grid[batch, 1, y_{dst}, x_{dst}] \\ output[batch, channel, y_{dst}, x_{dst}] = G(data[batch, channel, y_{src}, x_{src})\end{split}\]

\(x_{dst}\), \(y_{dst}\) enumerate all spatial locations in \(output\), and \(G()\) denotes the bilinear interpolation kernel. The out-boundary points will be padded with zeros.The shape of the output will be (data.shape[0], data.shape[1], grid.shape[2], grid.shape[3]).

The operator assumes that \(data\) has ‘NCHW’ layout and \(grid\) has been normalized to [-1, 1].

BilinearSampler often cooperates with GridGenerator which generates sampling grids for BilinearSampler. GridGenerator supports two kinds of transformation: affine and warp. If users want to design a CustomOp to manipulate \(grid\), please firstly refer to the code of GridGenerator.

Example 1:

## Zoom out data two times
data = array([[[[1, 4, 3, 6],
                [1, 8, 8, 9],
                [0, 4, 1, 5],
                [1, 0, 1, 3]]]])

affine_matrix = array([[2, 0, 0],
                       [0, 2, 0]])

affine_matrix = reshape(affine_matrix, shape=(1, 6))

grid = GridGenerator(data=affine_matrix, transform_type='affine', target_shape=(4, 4))

out = BilinearSampler(data, grid)

out
[[[[ 0,   0,     0,   0],
   [ 0,   3.5,   6.5, 0],
   [ 0,   1.25,  2.5, 0],
   [ 0,   0,     0,   0]]]

Example 2:

## shift data horizontally by -1 pixel

data = array([[[[1, 4, 3, 6],
                [1, 8, 8, 9],
                [0, 4, 1, 5],
                [1, 0, 1, 3]]]])

warp_maxtrix = array([[[[1, 1, 1, 1],
                        [1, 1, 1, 1],
                        [1, 1, 1, 1],
                        [1, 1, 1, 1]],
                       [[0, 0, 0, 0],
                        [0, 0, 0, 0],
                        [0, 0, 0, 0],
                        [0, 0, 0, 0]]]])

grid = GridGenerator(data=warp_matrix, transform_type='warp')
out = BilinearSampler(data, grid)

out
[[[[ 4,  3,  6,  0],
   [ 8,  8,  9,  0],
   [ 4,  1,  5,  0],
   [ 0,  1,  3,  0]]]

Defined in src/operator/bilinear_sampler.cc:L256

Parameters:
  • data (NDArray) – Input data to the BilinearsamplerOp.
  • grid (NDArray) – Input grid to the BilinearsamplerOp.grid has two channels: x_src, y_src
  • cudnn_off (boolean or None, optional, default=None) – whether to turn cudnn off
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.BlockGrad(data=None, out=None, name=None, **kwargs)

Stops gradient computation.

Stops the accumulated gradient of the inputs from flowing through this operator in the backward direction. In other words, this operator prevents the contribution of its inputs to be taken into account for computing gradients.

Example:

v1 = [1, 2]
v2 = [0, 1]
a = Variable('a')
b = Variable('b')
b_stop_grad = stop_gradient(3 * b)
loss = MakeLoss(b_stop_grad + a)

executor = loss.simple_bind(ctx=cpu(), a=(1,2), b=(1,2))
executor.forward(is_train=True, a=v1, b=v2)
executor.outputs
[ 1.  5.]

executor.backward()
executor.grad_arrays
[ 0.  0.]
[ 1.  1.]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L299

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.CTCLoss(data=None, label=None, data_lengths=None, label_lengths=None, use_data_lengths=_Null, use_label_lengths=_Null, blank_label=_Null, out=None, name=None, **kwargs)

Connectionist Temporal Classification Loss.

Note

The existing alias contrib_CTCLoss is deprecated.

The shapes of the inputs and outputs:

  • data: (sequence_length, batch_size, alphabet_size)
  • label: (batch_size, label_sequence_length)
  • out: (batch_size)

The data tensor consists of sequences of activation vectors (without applying softmax), with i-th channel in the last dimension corresponding to i-th label for i between 0 and alphabet_size-1 (i.e always 0-indexed). Alphabet size should include one additional value reserved for blank label. When blank_label is "first", the 0-th channel is be reserved for activation of blank label, or otherwise if it is “last”, (alphabet_size-1)-th channel should be reserved for blank label.

label is an index matrix of integers. When blank_label is "first", the value 0 is then reserved for blank label, and should not be passed in this matrix. Otherwise, when blank_label is "last", the value (alphabet_size-1) is reserved for blank label.

If a sequence of labels is shorter than label_sequence_length, use the special padding value at the end of the sequence to conform it to the correct length. The padding value is 0 when blank_label is "first", and -1 otherwise.

For example, suppose the vocabulary is [a, b, c], and in one batch we have three sequences ‘ba’, ‘cbb’, and ‘abac’. When blank_label is "first", we can index the labels as {‘a’: 1, ‘b’: 2, ‘c’: 3}, and we reserve the 0-th channel for blank label in data tensor. The resulting label tensor should be padded to be:

[[2, 1, 0, 0], [3, 2, 2, 0], [1, 2, 1, 3]]

When blank_label is "last", we can index the labels as {‘a’: 0, ‘b’: 1, ‘c’: 2}, and we reserve the channel index 3 for blank label in data tensor. The resulting label tensor should be padded to be:

[[1, 0, -1, -1], [2, 1, 1, -1], [0, 1, 0, 2]]

out is a list of CTC loss values, one per example in the batch.

See Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks, A. Graves et al. for more information on the definition and the algorithm.

Defined in src/operator/nn/ctc_loss.cc:L100

Parameters:
  • data (NDArray) – Input ndarray
  • label (NDArray) – Ground-truth labels for the loss.
  • data_lengths (NDArray) – Lengths of data for each of the samples. Only required when use_data_lengths is true.
  • label_lengths (NDArray) – Lengths of labels for each of the samples. Only required when use_label_lengths is true.
  • use_data_lengths (boolean, optional, default=0) – Whether the data lenghts are decided by data_lengths. If false, the lengths are equal to the max sequence length.
  • use_label_lengths (boolean, optional, default=0) – Whether the label lenghts are decided by label_lengths, or derived from padding_mask. If false, the lengths are derived from the first occurrence of the value of padding_mask. The value of padding_mask is 0 when first CTC label is reserved for blank, and -1 when last label is reserved for blank. See blank_label.
  • blank_label ({'first', 'last'},optional, default='first') – Set the label that is reserved for blank label.If “first”, 0-th label is reserved, and label values for tokens in the vocabulary are between 1 and alphabet_size-1, and the padding mask is -1. If “last”, last label value alphabet_size-1 is reserved for blank label instead, and label values for tokens in the vocabulary are between 0 and alphabet_size-2, and the padding mask is 0.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Cast(data=None, dtype=_Null, out=None, name=None, **kwargs)

Casts all elements of the input to a new type.

Note

Cast is deprecated. Use cast instead.

Example:

cast([0.9, 1.3], dtype='int32') = [0, 1]
cast([1e20, 11.1], dtype='float16') = [inf, 11.09375]
cast([300, 11.1, 10.9, -1, -3], dtype='uint8') = [44, 11, 10, 255, 253]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L634

Parameters:
  • data (NDArray) – The input.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8'}, required) – Output data type.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Concat(*data, **kwargs)

Joins input arrays along a given axis.

Note

Concat is deprecated. Use concat instead.

The dimensions of the input arrays should be the same except the axis along which they will be concatenated. The dimension of the output array along the concatenated axis will be equal to the sum of the corresponding dimensions of the input arrays.

The storage type of concat output depends on storage types of inputs

  • concat(csr, csr, ..., csr, dim=0) = csr
  • otherwise, concat generates output with default storage

Example:

x = [[1,1],[2,2]]
y = [[3,3],[4,4],[5,5]]
z = [[6,6], [7,7],[8,8]]

concat(x,y,z,dim=0) = [[ 1.,  1.],
                       [ 2.,  2.],
                       [ 3.,  3.],
                       [ 4.,  4.],
                       [ 5.,  5.],
                       [ 6.,  6.],
                       [ 7.,  7.],
                       [ 8.,  8.]]

Note that you cannot concat x,y,z along dimension 1 since dimension
0 is not the same for all the input arrays.

concat(y,z,dim=1) = [[ 3.,  3.,  6.,  6.],
                      [ 4.,  4.,  7.,  7.],
                      [ 5.,  5.,  8.,  8.]]

Defined in src/operator/nn/concat.cc:L371

Parameters:
  • data (NDArray[]) – List of arrays to concatenate
  • dim (int, optional, default='1') – the dimension to be concated.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Convolution(data=None, weight=None, bias=None, kernel=_Null, stride=_Null, dilate=_Null, pad=_Null, num_filter=_Null, num_group=_Null, workspace=_Null, no_bias=_Null, cudnn_tune=_Null, cudnn_off=_Null, layout=_Null, out=None, name=None, **kwargs)

Compute N-D convolution on (N+2)-D input.

In the 2-D convolution, given input data with shape (batch_size, channel, height, width), the output is computed by

\[out[n,i,:,:] = bias[i] + \sum_{j=0}^{channel} data[n,j,:,:] \star weight[i,j,:,:]\]

where \(\star\) is the 2-D cross-correlation operator.

For general 2-D convolution, the shapes are

  • data: (batch_size, channel, height, width)
  • weight: (num_filter, channel, kernel[0], kernel[1])
  • bias: (num_filter,)
  • out: (batch_size, num_filter, out_height, out_width).

Define:

f(x,k,p,s,d) = floor((x+2*p-d*(k-1)-1)/s)+1

then we have:

out_height=f(height, kernel[0], pad[0], stride[0], dilate[0])
out_width=f(width, kernel[1], pad[1], stride[1], dilate[1])

If no_bias is set to be true, then the bias term is ignored.

The default data layout is NCHW, namely (batch_size, channel, height, width). We can choose other layouts such as NWC.

If num_group is larger than 1, denoted by g, then split the input data evenly into g parts along the channel axis, and also evenly split weight along the first dimension. Next compute the convolution on the i-th part of the data with the i-th weight part. The output is obtained by concatenating all the g results.

1-D convolution does not have height dimension but only width in space.

  • data: (batch_size, channel, width)
  • weight: (num_filter, channel, kernel[0])
  • bias: (num_filter,)
  • out: (batch_size, num_filter, out_width).

3-D convolution adds an additional depth dimension besides height and width. The shapes are

  • data: (batch_size, channel, depth, height, width)
  • weight: (num_filter, channel, kernel[0], kernel[1], kernel[2])
  • bias: (num_filter,)
  • out: (batch_size, num_filter, out_depth, out_height, out_width).

Both weight and bias are learnable parameters.

There are other options to tune the performance.

  • cudnn_tune: enable this option leads to higher startup time but may give faster speed. Options are
    • off: no tuning
    • limited_workspace:run test and pick the fastest algorithm that doesn’t exceed workspace limit.
    • fastest: pick the fastest algorithm and ignore workspace limit.
    • None (default): the behavior is determined by environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT. 0 for off, 1 for limited workspace (default), 2 for fastest.
  • workspace: A large number leads to more (GPU) memory usage but may improve the performance.

Defined in src/operator/nn/convolution.cc:L472

Parameters:
  • data (NDArray) – Input data to the ConvolutionOp.
  • weight (NDArray) – Weight matrix.
  • bias (NDArray) – Bias parameter.
  • kernel (Shape(tuple), required) – Convolution kernel size: (w,), (h, w) or (d, h, w)
  • stride (Shape(tuple), optional, default=[]) – Convolution stride: (w,), (h, w) or (d, h, w). Defaults to 1 for each dimension.
  • dilate (Shape(tuple), optional, default=[]) – Convolution dilate: (w,), (h, w) or (d, h, w). Defaults to 1 for each dimension.
  • pad (Shape(tuple), optional, default=[]) – Zero pad for convolution: (w,), (h, w) or (d, h, w). Defaults to no padding.
  • num_filter (int (non-negative), required) – Convolution filter(channel) number
  • num_group (int (non-negative), optional, default=1) – Number of group partitions.
  • workspace (long (non-negative), optional, default=1024) – Maximum temporary workspace allowed (MB) in convolution.This parameter has two usages. When CUDNN is not used, it determines the effective batch size of the convolution kernel. When CUDNN is used, it controls the maximum temporary storage used for tuning the best CUDNN kernel when limited_workspace strategy is used.
  • no_bias (boolean, optional, default=0) – Whether to disable bias parameter.
  • cudnn_tune ({None, 'fastest', 'limited_workspace', 'off'},optional, default='None') – Whether to pick convolution algo by running performance test.
  • cudnn_off (boolean, optional, default=0) – Turn off cudnn for this layer.
  • layout ({None, 'NCDHW', 'NCHW', 'NCW', 'NDHWC', 'NHWC'},optional, default='None') – Set layout for input, output and weight. Empty for default layout: NCW for 1d, NCHW for 2d and NCDHW for 3d.NHWC and NDHWC are only supported on GPU.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Convolution_v1(data=None, weight=None, bias=None, kernel=_Null, stride=_Null, dilate=_Null, pad=_Null, num_filter=_Null, num_group=_Null, workspace=_Null, no_bias=_Null, cudnn_tune=_Null, cudnn_off=_Null, layout=_Null, out=None, name=None, **kwargs)

This operator is DEPRECATED. Apply convolution to input then add a bias.

Parameters:
  • data (NDArray) – Input data to the ConvolutionV1Op.
  • weight (NDArray) – Weight matrix.
  • bias (NDArray) – Bias parameter.
  • kernel (Shape(tuple), required) – convolution kernel size: (h, w) or (d, h, w)
  • stride (Shape(tuple), optional, default=[]) – convolution stride: (h, w) or (d, h, w)
  • dilate (Shape(tuple), optional, default=[]) – convolution dilate: (h, w) or (d, h, w)
  • pad (Shape(tuple), optional, default=[]) – pad for convolution: (h, w) or (d, h, w)
  • num_filter (int (non-negative), required) – convolution filter(channel) number
  • num_group (int (non-negative), optional, default=1) – Number of group partitions. Equivalent to slicing input into num_group partitions, apply convolution on each, then concatenate the results
  • workspace (long (non-negative), optional, default=1024) – Maximum temporary workspace allowed for convolution (MB).This parameter determines the effective batch size of the convolution kernel, which may be smaller than the given batch size. Also, the workspace will be automatically enlarged to make sure that we can run the kernel with batch_size=1
  • no_bias (boolean, optional, default=0) – Whether to disable bias parameter.
  • cudnn_tune ({None, 'fastest', 'limited_workspace', 'off'},optional, default='None') – Whether to pick convolution algo by running performance test. Leads to higher startup time but may give faster speed. Options are: ‘off’: no tuning ‘limited_workspace’: run test and pick the fastest algorithm that doesn’t exceed workspace limit. ‘fastest’: pick the fastest algorithm and ignore workspace limit. If set to None (default), behavior is determined by environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT: 0 for off, 1 for limited workspace (default), 2 for fastest.
  • cudnn_off (boolean, optional, default=0) – Turn off cudnn for this layer.
  • layout ({None, 'NCDHW', 'NCHW', 'NDHWC', 'NHWC'},optional, default='None') – Set layout for input, output and weight. Empty for default layout: NCHW for 2d and NCDHW for 3d.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Correlation(data1=None, data2=None, kernel_size=_Null, max_displacement=_Null, stride1=_Null, stride2=_Null, pad_size=_Null, is_multiply=_Null, out=None, name=None, **kwargs)

Applies correlation to inputs.

The correlation layer performs multiplicative patch comparisons between two feature maps.

Given two multi-channel feature maps \(f_{1}, f_{2}\), with \(w\), \(h\), and \(c\) being their width, height, and number of channels, the correlation layer lets the network compare each patch from \(f_{1}\) with each patch from \(f_{2}\).

For now we consider only a single comparison of two patches. The ‘correlation’ of two patches centered at \(x_{1}\) in the first map and \(x_{2}\) in the second map is then defined as:

\[c(x_{1}, x_{2}) = \sum_{o \in [-k,k] \times [-k,k]} \]

for a square patch of size \(K:=2k+1\).

Note that the equation above is identical to one step of a convolution in neural networks, but instead of convolving data with a filter, it convolves data with other data. For this reason, it has no training weights.

Computing \(c(x_{1}, x_{2})\) involves \(c * K^{2}\) multiplications. Comparing all patch combinations involves \(w^{2}*h^{2}\) such computations.

Given a maximum displacement \(d\), for each location \(x_{1}\) it computes correlations \(c(x_{1}, x_{2})\) only in a neighborhood of size \(D:=2d+1\), by limiting the range of \(x_{2}\). We use strides \(s_{1}, s_{2}\), to quantize \(x_{1}\) globally and to quantize \(x_{2}\) within the neighborhood centered around \(x_{1}\).

The final output is defined by the following expression:

\[out[n, q, i, j] = c(x_{i, j}, x_{q})\]

where \(i\) and \(j\) enumerate spatial locations in \(f_{1}\), and \(q\) denotes the \(q^{th}\) neighborhood of \(x_{i,j}\).

Defined in src/operator/correlation.cc:L198

Parameters:
  • data1 (NDArray) – Input data1 to the correlation.
  • data2 (NDArray) – Input data2 to the correlation.
  • kernel_size (int (non-negative), optional, default=1) – kernel size for Correlation must be an odd number
  • max_displacement (int (non-negative), optional, default=1) – Max displacement of Correlation
  • stride1 (int (non-negative), optional, default=1) – stride1 quantize data1 globally
  • stride2 (int (non-negative), optional, default=1) – stride2 quantize data2 within the neighborhood centered around data1
  • pad_size (int (non-negative), optional, default=0) – pad for Correlation
  • is_multiply (boolean, optional, default=1) – operation type is either multiplication or subduction
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Crop(*data, **kwargs)

Note

Crop is deprecated. Use slice instead.

Crop the 2nd and 3rd dim of input data, with the corresponding size of h_w or with width and height of the second input symbol, i.e., with one input, we need h_w to specify the crop height and width, otherwise the second input symbol’s size will be used

Defined in src/operator/crop.cc:L50

Parameters:
  • data (Symbol or Symbol[]) – Tensor or List of Tensors, the second input will be used as crop_like shape reference
  • offset (Shape(tuple), optional, default=[0,0]) – crop offset coordinate: (y, x)
  • h_w (Shape(tuple), optional, default=[0,0]) – crop height and width: (h, w)
  • center_crop (boolean, optional, default=0) – If set to true, then it will use be the center_crop,or it will crop using the shape of crop_like
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Custom(*data, **kwargs)

Apply a custom operator implemented in a frontend language (like Python).

Custom operators should override required methods like forward and backward. The custom operator must be registered before it can be used. Please check the tutorial here: /faq/new_op.html.

Defined in src/operator/custom/custom.cc:L546

Parameters:
  • data (NDArray[]) – Input data for the custom operator.
  • op_type (string) – Name of the custom operator. This is the name that is passed to mx.operator.register to register the operator.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

Example

Applies a custom operator named my_custom_operator to input.

>>> output = mx.symbol.Custom(op_type='my_custom_operator', data=input)
mxnet.ndarray.Deconvolution(data=None, weight=None, bias=None, kernel=_Null, stride=_Null, dilate=_Null, pad=_Null, adj=_Null, target_shape=_Null, num_filter=_Null, num_group=_Null, workspace=_Null, no_bias=_Null, cudnn_tune=_Null, cudnn_off=_Null, layout=_Null, out=None, name=None, **kwargs)

Computes 1D or 2D transposed convolution (aka fractionally strided convolution) of the input tensor. This operation can be seen as the gradient of Convolution operation with respect to its input. Convolution usually reduces the size of the input. Transposed convolution works the other way, going from a smaller input to a larger output while preserving the connectivity pattern.

Parameters:
  • data (NDArray) – Input tensor to the deconvolution operation.
  • weight (NDArray) – Weights representing the kernel.
  • bias (NDArray) – Bias added to the result after the deconvolution operation.
  • kernel (Shape(tuple), required) – Deconvolution kernel size: (w,), (h, w) or (d, h, w). This is same as the kernel size used for the corresponding convolution
  • stride (Shape(tuple), optional, default=[]) – The stride used for the corresponding convolution: (w,), (h, w) or (d, h, w). Defaults to 1 for each dimension.
  • dilate (Shape(tuple), optional, default=[]) – Dilation factor for each dimension of the input: (w,), (h, w) or (d, h, w). Defaults to 1 for each dimension.
  • pad (Shape(tuple), optional, default=[]) – The amount of implicit zero padding added during convolution for each dimension of the input: (w,), (h, w) or (d, h, w). (kernel-1)/2 is usually a good choice. If target_shape is set, pad will be ignored and a padding that will generate the target shape will be used. Defaults to no padding.
  • adj (Shape(tuple), optional, default=[]) – Adjustment for output shape: (w,), (h, w) or (d, h, w). If target_shape is set, adj will be ignored and computed accordingly.
  • target_shape (Shape(tuple), optional, default=[]) – Shape of the output tensor: (w,), (h, w) or (d, h, w).
  • num_filter (int (non-negative), required) – Number of output filters.
  • num_group (int (non-negative), optional, default=1) – Number of groups partition.
  • workspace (long (non-negative), optional, default=512) – Maximum temporary workspace allowed (MB) in deconvolution.This parameter has two usages. When CUDNN is not used, it determines the effective batch size of the deconvolution kernel. When CUDNN is used, it controls the maximum temporary storage used for tuning the best CUDNN kernel when limited_workspace strategy is used.
  • no_bias (boolean, optional, default=1) – Whether to disable bias parameter.
  • cudnn_tune ({None, 'fastest', 'limited_workspace', 'off'},optional, default='None') – Whether to pick convolution algorithm by running performance test.
  • cudnn_off (boolean, optional, default=0) – Turn off cudnn for this layer.
  • layout ({None, 'NCDHW', 'NCHW', 'NCW', 'NDHWC', 'NHWC'},optional, default='None') – Set layout for input, output and weight. Empty for default layout, NCW for 1d, NCHW for 2d and NCDHW for 3d.NHWC and NDHWC are only supported on GPU.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Dropout(data=None, p=_Null, mode=_Null, axes=_Null, cudnn_off=_Null, out=None, name=None, **kwargs)

Applies dropout operation to input array.

  • During training, each element of the input is set to zero with probability p. The whole array is rescaled by \(1/(1-p)\) to keep the expected sum of the input unchanged.
  • During testing, this operator does not change the input if mode is ‘training’. If mode is ‘always’, the same computaion as during training will be applied.

Example:

random.seed(998)
input_array = array([[3., 0.5,  -0.5,  2., 7.],
                    [2., -0.4,   7.,  3., 0.2]])
a = symbol.Variable('a')
dropout = symbol.Dropout(a, p = 0.2)
executor = dropout.simple_bind(a = input_array.shape)

## If training
executor.forward(is_train = True, a = input_array)
executor.outputs
[[ 3.75   0.625 -0.     2.5    8.75 ]
 [ 2.5   -0.5    8.75   3.75   0.   ]]

## If testing
executor.forward(is_train = False, a = input_array)
executor.outputs
[[ 3.     0.5   -0.5    2.     7.   ]
 [ 2.    -0.4    7.     3.     0.2  ]]

Defined in src/operator/nn/dropout.cc:L95

Parameters:
  • data (NDArray) – Input array to which dropout will be applied.
  • p (float, optional, default=0.5) – Fraction of the input that gets dropped out during training time.
  • mode ({'always', 'training'},optional, default='training') – Whether to only turn on dropout during training or to also turn on for inference.
  • axes (Shape(tuple), optional, default=[]) – Axes for variational dropout kernel.
  • cudnn_off (boolean or None, optional, default=0) – Whether to turn off cudnn in dropout operator. This option is ignored if axes is specified.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ElementWiseSum(*args, **kwargs)

Adds all input arguments element-wise.

\[add\_n(a_1, a_2, ..., a_n) = a_1 + a_2 + ... + a_n\]

add_n is potentially more efficient than calling add by n times.

The storage type of add_n output depends on storage types of inputs

  • add_n(row_sparse, row_sparse, ..) = row_sparse
  • add_n(default, csr, default) = default
  • add_n(any input combinations longer than 4 (>4) with at least one default type) = default
  • otherwise, add_n falls all inputs back to default storage and generates default storage

Defined in src/operator/tensor/elemwise_sum.cc:L155

Parameters:
  • args (NDArray[]) – Positional input arguments
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Embedding(data=None, weight=None, input_dim=_Null, output_dim=_Null, dtype=_Null, sparse_grad=_Null, out=None, name=None, **kwargs)

Maps integer indices to vector representations (embeddings).

This operator maps words to real-valued vectors in a high-dimensional space, called word embeddings. These embeddings can capture semantic and syntactic properties of the words. For example, it has been noted that in the learned embedding spaces, similar words tend to be close to each other and dissimilar words far apart.

For an input array of shape (d1, ..., dK), the shape of an output array is (d1, ..., dK, output_dim). All the input values should be integers in the range [0, input_dim).

If the input_dim is ip0 and output_dim is op0, then shape of the embedding weight matrix must be (ip0, op0).

By default, if any index mentioned is too large, it is replaced by the index that addresses the last vector in an embedding matrix.

Examples:

input_dim = 4
output_dim = 5

// Each row in weight matrix y represents a word. So, y = (w0,w1,w2,w3)
y = [[  0.,   1.,   2.,   3.,   4.],
     [  5.,   6.,   7.,   8.,   9.],
     [ 10.,  11.,  12.,  13.,  14.],
     [ 15.,  16.,  17.,  18.,  19.]]

// Input array x represents n-grams(2-gram). So, x = [(w1,w3), (w0,w2)]
x = [[ 1.,  3.],
     [ 0.,  2.]]

// Mapped input x to its vector representation y.
Embedding(x, y, 4, 5) = [[[  5.,   6.,   7.,   8.,   9.],
                          [ 15.,  16.,  17.,  18.,  19.]],

                         [[  0.,   1.,   2.,   3.,   4.],
                          [ 10.,  11.,  12.,  13.,  14.]]]

The storage type of weight can be either row_sparse or default.

Note

If “sparse_grad” is set to True, the storage type of gradient w.r.t weights will be “row_sparse”. Only a subset of optimizers support sparse gradients, including SGD, AdaGrad and Adam. Note that by default lazy updates is turned on, which may perform differently from standard updates. For more details, please check the Optimization API at: /api/python/optimization/optimization.html

Defined in src/operator/tensor/indexing_op.cc:L519

Parameters:
  • data (NDArray) – The input array to the embedding operator.
  • weight (NDArray) – The embedding weight matrix.
  • input_dim (int, required) – Vocabulary size of the input indices.
  • output_dim (int, required) – Dimension of the embedding vectors.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8'},optional, default='float32') – Data type of weight.
  • sparse_grad (boolean, optional, default=0) – Compute row sparse gradient in the backward calculation. If set to True, the grad’s storage type is row_sparse.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Flatten(data=None, out=None, name=None, **kwargs)

Flattens the input array into a 2-D array by collapsing the higher dimensions.

Note

Flatten is deprecated. Use flatten instead.

For an input array with shape (d1, d2, ..., dk), flatten operation reshapes the input array into an output array of shape (d1, d2*...*dk).

Note that the bahavior of this function is different from numpy.ndarray.flatten, which behaves similar to mxnet.ndarray.reshape((-1,)).

Example:

x = [[
    [1,2,3],
    [4,5,6],
    [7,8,9]
],
[    [1,2,3],
    [4,5,6],
    [7,8,9]
]],

flatten(x) = [[ 1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9.],
   [ 1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9.]]

Defined in src/operator/tensor/matrix_op.cc:L291

Parameters:
  • data (NDArray) – Input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.FullyConnected(data=None, weight=None, bias=None, num_hidden=_Null, no_bias=_Null, flatten=_Null, out=None, name=None, **kwargs)

Applies a linear transformation: \(Y = XW^T + b\).

If flatten is set to be true, then the shapes are:

  • data: (batch_size, x1, x2, ..., xn)
  • weight: (num_hidden, x1 * x2 * ... * xn)
  • bias: (num_hidden,)
  • out: (batch_size, num_hidden)

If flatten is set to be false, then the shapes are:

  • data: (x1, x2, ..., xn, input_dim)
  • weight: (num_hidden, input_dim)
  • bias: (num_hidden,)
  • out: (x1, x2, ..., xn, num_hidden)

The learnable parameters include both weight and bias.

If no_bias is set to be true, then the bias term is ignored.

Note

The sparse support for FullyConnected is limited to forward evaluation with row_sparse weight and bias, where the length of weight.indices and bias.indices must be equal to num_hidden. This could be useful for model inference with row_sparse weights trained with importance sampling or noise contrastive estimation.

To compute linear transformation with ‘csr’ sparse data, sparse.dot is recommended instead of sparse.FullyConnected.

Defined in src/operator/nn/fully_connected.cc:L277

Parameters:
  • data (NDArray) – Input data.
  • weight (NDArray) – Weight matrix.
  • bias (NDArray) – Bias parameter.
  • num_hidden (int, required) – Number of hidden nodes of the output.
  • no_bias (boolean, optional, default=0) – Whether to disable bias parameter.
  • flatten (boolean, optional, default=1) – Whether to collapse all but the first axis of the input data tensor.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.GridGenerator(data=None, transform_type=_Null, target_shape=_Null, out=None, name=None, **kwargs)

Generates 2D sampling grid for bilinear sampling.

Parameters:
  • data (NDArray) – Input data to the function.
  • transform_type ({'affine', 'warp'}, required) – The type of transformation. For affine, input data should be an affine matrix of size (batch, 6). For warp, input data should be an optical flow of size (batch, 2, h, w).
  • target_shape (Shape(tuple), optional, default=[0,0]) – Specifies the output shape (H, W). This is required if transformation type is affine. If transformation type is warp, this parameter is ignored.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.IdentityAttachKLSparseReg(data=None, sparseness_target=_Null, penalty=_Null, momentum=_Null, out=None, name=None, **kwargs)

Apply a sparse regularization to the output a sigmoid activation function.

Parameters:
  • data (NDArray) – Input data.
  • sparseness_target (float, optional, default=0.100000001) – The sparseness target
  • penalty (float, optional, default=0.00100000005) – The tradeoff parameter for the sparseness penalty
  • momentum (float, optional, default=0.899999976) – The momentum for running average
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.InstanceNorm(data=None, gamma=None, beta=None, eps=_Null, out=None, name=None, **kwargs)

Applies instance normalization to the n-dimensional input array.

This operator takes an n-dimensional input array where (n>2) and normalizes the input using the following formula:

\[out = \frac{x - mean[data]}{ \sqrt{Var[data]} + \epsilon} * gamma + beta\]

This layer is similar to batch normalization layer (BatchNorm) with two differences: first, the normalization is carried out per example (instance), not over a batch. Second, the same normalization is applied both at test and train time. This operation is also known as contrast normalization.

If the input data is of shape [batch, channel, spacial_dim1, spacial_dim2, ...], gamma and beta parameters must be vectors of shape [channel].

This implementation is based on paper:

[1]Instance Normalization: The Missing Ingredient for Fast Stylization, D. Ulyanov, A. Vedaldi, V. Lempitsky, 2016 (arXiv:1607.08022v2).

Examples:

// Input of shape (2,1,2)
x = [[[ 1.1,  2.2]],
     [[ 3.3,  4.4]]]

// gamma parameter of length 1
gamma = [1.5]

// beta parameter of length 1
beta = [0.5]

// Instance normalization is calculated with the above formula
InstanceNorm(x,gamma,beta) = [[[-0.997527  ,  1.99752665]],
                              [[-0.99752653,  1.99752724]]]

Defined in src/operator/instance_norm.cc:L95

Parameters:
  • data (NDArray) – An n-dimensional input array (n > 2) of the form [batch, channel, spatial_dim1, spatial_dim2, ...].
  • gamma (NDArray) – A vector of length ‘channel’, which multiplies the normalized input.
  • beta (NDArray) – A vector of length ‘channel’, which is added to the product of the normalized input and the weight.
  • eps (float, optional, default=0.00100000005) – An epsilon parameter to prevent division by 0.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.L2Normalization(data=None, eps=_Null, mode=_Null, out=None, name=None, **kwargs)

Normalize the input array using the L2 norm.

For 1-D NDArray, it computes:

out = data / sqrt(sum(data ** 2) + eps)

For N-D NDArray, if the input array has shape (N, N, ..., N),

with mode = instance, it normalizes each instance in the multidimensional array by its L2 norm.:

for i in 0...N
  out[i,:,:,...,:] = data[i,:,:,...,:] / sqrt(sum(data[i,:,:,...,:] ** 2) + eps)

with mode = channel, it normalizes each channel in the array by its L2 norm.:

for i in 0...N
  out[:,i,:,...,:] = data[:,i,:,...,:] / sqrt(sum(data[:,i,:,...,:] ** 2) + eps)

with mode = spatial, it normalizes the cross channel norm for each position in the array by its L2 norm.:

for dim in 2...N
  for i in 0...N
    out[.....,i,...] = take(out, indices=i, axis=dim) / sqrt(sum(take(out, indices=i, axis=dim) ** 2) + eps)
        -dim-

Example:

x = [[[1,2],
      [3,4]],
     [[2,2],
      [5,6]]]

L2Normalization(x, mode='instance')
=[[[ 0.18257418  0.36514837]
   [ 0.54772252  0.73029673]]
  [[ 0.24077171  0.24077171]
   [ 0.60192931  0.72231513]]]

L2Normalization(x, mode='channel')
=[[[ 0.31622776  0.44721359]
   [ 0.94868326  0.89442718]]
  [[ 0.37139067  0.31622776]
   [ 0.92847669  0.94868326]]]

L2Normalization(x, mode='spatial')
=[[[ 0.44721359  0.89442718]
   [ 0.60000002  0.80000001]]
  [[ 0.70710677  0.70710677]
   [ 0.6401844   0.76822126]]]

Defined in src/operator/l2_normalization.cc:L196

Parameters:
  • data (NDArray) – Input array to normalize.
  • eps (float, optional, default=1.00000001e-10) – A small constant for numerical stability.
  • mode ({'channel', 'instance', 'spatial'},optional, default='instance') – Specify the dimension along which to compute L2 norm.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.LRN(data=None, alpha=_Null, beta=_Null, knorm=_Null, nsize=_Null, out=None, name=None, **kwargs)

Applies local response normalization to the input.

The local response normalization layer performs “lateral inhibition” by normalizing over local input regions.

If \(a_{x,y}^{i}\) is the activity of a neuron computed by applying kernel \(i\) at position \((x, y)\) and then applying the ReLU nonlinearity, the response-normalized activity \(b_{x,y}^{i}\) is given by the expression:

\[b_{x,y}^{i} = \frac{a_{x,y}^{i}}{\Bigg({k + \frac{\alpha}{n} \sum_{j=max(0, i-\frac{n}{2})}^{min(N-1, i+\frac{n}{2})} (a_{x,y}^{j})^{2}}\Bigg)^{\beta}}\]

where the sum runs over \(n\) “adjacent” kernel maps at the same spatial position, and \(N\) is the total number of kernels in the layer.

Defined in src/operator/nn/lrn.cc:L164

Parameters:
  • data (NDArray) – Input data to LRN
  • alpha (float, optional, default=9.99999975e-05) – The variance scaling parameter \(lpha\) in the LRN expression.
  • beta (float, optional, default=0.75) – The power parameter \(eta\) in the LRN expression.
  • knorm (float, optional, default=2) – The parameter \(k\) in the LRN expression.
  • nsize (int (non-negative), required) – normalization window width in elements.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.LayerNorm(data=None, gamma=None, beta=None, axis=_Null, eps=_Null, output_mean_var=_Null, out=None, name=None, **kwargs)

Layer normalization.

Normalizes the channels of the input tensor by mean and variance, and applies a scale gamma as well as offset beta.

Assume the input has more than one dimension and we normalize along axis 1. We first compute the mean and variance along this axis and then compute the normalized output, which has the same shape as input, as following:

\[out = \frac{data - mean(data, axis)}{\sqrt{var(data, axis) + \epsilon}} * gamma + beta\]

Both gamma and beta are learnable parameters.

Unlike BatchNorm and InstanceNorm, the mean and var are computed along the channel dimension.

Assume the input has size k on axis 1, then both gamma and beta have shape (k,). If output_mean_var is set to be true, then outputs both data_mean and data_std. Note that no gradient will be passed through these two outputs.

The parameter axis specifies which axis of the input shape denotes the ‘channel’ (separately normalized groups). The default is -1, which sets the channel axis to be the last item in the input shape.

Defined in src/operator/nn/layer_norm.cc:L155

Parameters:
  • data (NDArray) – Input data to layer normalization
  • gamma (NDArray) – gamma array
  • beta (NDArray) – beta array
  • axis (int, optional, default='-1') – The axis to perform layer normalization. Usually, this should be be axis of the channel dimension. Negative values means indexing from right to left.
  • eps (float, optional, default=9.99999975e-06) – An epsilon parameter to prevent division by 0.
  • output_mean_var (boolean, optional, default=0) – Output the mean and std calculated along the given axis.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.LeakyReLU(data=None, gamma=None, act_type=_Null, slope=_Null, lower_bound=_Null, upper_bound=_Null, out=None, name=None, **kwargs)

Applies Leaky rectified linear unit activation element-wise to the input.

Leaky ReLUs attempt to fix the “dying ReLU” problem by allowing a small slope when the input is negative and has a slope of one when input is positive.

The following modified ReLU Activation functions are supported:

  • elu: Exponential Linear Unit. y = x > 0 ? x : slope * (exp(x)-1)
  • selu: Scaled Exponential Linear Unit. y = lambda * (x > 0 ? x : alpha * (exp(x) - 1)) where lambda = 1.0507009873554804934193349852946 and alpha = 1.6732632423543772848170429916717.
  • leaky: Leaky ReLU. y = x > 0 ? x : slope * x
  • prelu: Parametric ReLU. This is same as leaky except that slope is learnt during training.
  • rrelu: Randomized ReLU. same as leaky but the slope is uniformly and randomly chosen from [lower_bound, upper_bound) for training, while fixed to be (lower_bound+upper_bound)/2 for inference.

Defined in src/operator/leaky_relu.cc:L65

Parameters:
  • data (NDArray) – Input data to activation function.
  • gamma (NDArray) – Slope parameter for PReLU. Only required when act_type is ‘prelu’. It should be either a vector of size 1, or the same size as the second dimension of data.
  • act_type ({'elu', 'gelu', 'leaky', 'prelu', 'rrelu', 'selu'},optional, default='leaky') – Activation function to be applied.
  • slope (float, optional, default=0.25) – Init slope for the activation. (For leaky and elu only)
  • lower_bound (float, optional, default=0.125) – Lower bound of random slope. (For rrelu only)
  • upper_bound (float, optional, default=0.333999991) – Upper bound of random slope. (For rrelu only)
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.LinearRegressionOutput(data=None, label=None, grad_scale=_Null, out=None, name=None, **kwargs)

Computes and optimizes for squared loss during backward propagation. Just outputs data during forward propagation.

If \(\hat{y}_i\) is the predicted value of the i-th sample, and \(y_i\) is the corresponding target value, then the squared loss estimated over \(n\) samples is defined as

\(\text{SquaredLoss}(\textbf{Y}, \hat{\textbf{Y}} ) = \frac{1}{n} \sum_{i=0}^{n-1} \lVert \textbf{y}_i - \hat{\textbf{y}}_i \rVert_2\)

Note

Use the LinearRegressionOutput as the final output layer of a net.

The storage type of label can be default or csr

  • LinearRegressionOutput(default, default) = default
  • LinearRegressionOutput(default, csr) = default

By default, gradients of this loss function are scaled by factor 1/m, where m is the number of regression outputs of a training example. The parameter grad_scale can be used to change this scale to grad_scale/m.

Defined in src/operator/regression_output.cc:L92

Parameters:
  • data (NDArray) – Input data to the function.
  • label (NDArray) – Input label to the function.
  • grad_scale (float, optional, default=1) – Scale the gradient by a float factor
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.LogisticRegressionOutput(data=None, label=None, grad_scale=_Null, out=None, name=None, **kwargs)

Applies a logistic function to the input.

The logistic function, also known as the sigmoid function, is computed as \(\frac{1}{1+exp(-\textbf{x})}\).

Commonly, the sigmoid is used to squash the real-valued output of a linear model \(wTx+b\) into the [0,1] range so that it can be interpreted as a probability. It is suitable for binary classification or probability prediction tasks.

Note

Use the LogisticRegressionOutput as the final output layer of a net.

The storage type of label can be default or csr

  • LogisticRegressionOutput(default, default) = default
  • LogisticRegressionOutput(default, csr) = default

The loss function used is the Binary Cross Entropy Loss:

\(-{(y\log(p) + (1 - y)\log(1 - p))}\)

Where y is the ground truth probability of positive outcome for a given example, and p the probability predicted by the model. By default, gradients of this loss function are scaled by factor 1/m, where m is the number of regression outputs of a training example. The parameter grad_scale can be used to change this scale to grad_scale/m.

Defined in src/operator/regression_output.cc:L152

Parameters:
  • data (NDArray) – Input data to the function.
  • label (NDArray) – Input label to the function.
  • grad_scale (float, optional, default=1) – Scale the gradient by a float factor
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.MAERegressionOutput(data=None, label=None, grad_scale=_Null, out=None, name=None, **kwargs)

Computes mean absolute error of the input.

MAE is a risk metric corresponding to the expected value of the absolute error.

If \(\hat{y}_i\) is the predicted value of the i-th sample, and \(y_i\) is the corresponding target value, then the mean absolute error (MAE) estimated over \(n\) samples is defined as

\(\text{MAE}(\textbf{Y}, \hat{\textbf{Y}} ) = \frac{1}{n} \sum_{i=0}^{n-1} \lVert \textbf{y}_i - \hat{\textbf{y}}_i \rVert_1\)

Note

Use the MAERegressionOutput as the final output layer of a net.

The storage type of label can be default or csr

  • MAERegressionOutput(default, default) = default
  • MAERegressionOutput(default, csr) = default

By default, gradients of this loss function are scaled by factor 1/m, where m is the number of regression outputs of a training example. The parameter grad_scale can be used to change this scale to grad_scale/m.

Defined in src/operator/regression_output.cc:L120

Parameters:
  • data (NDArray) – Input data to the function.
  • label (NDArray) – Input label to the function.
  • grad_scale (float, optional, default=1) – Scale the gradient by a float factor
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.MakeLoss(data=None, grad_scale=_Null, valid_thresh=_Null, normalization=_Null, out=None, name=None, **kwargs)

Make your own loss function in network construction.

This operator accepts a customized loss function symbol as a terminal loss and the symbol should be an operator with no backward dependency. The output of this function is the gradient of loss with respect to the input data.

For example, if you are a making a cross entropy loss function. Assume out is the predicted output and label is the true label, then the cross entropy can be defined as:

cross_entropy = label * log(out) + (1 - label) * log(1 - out)
loss = MakeLoss(cross_entropy)

We will need to use MakeLoss when we are creating our own loss function or we want to combine multiple loss functions. Also we may want to stop some variables’ gradients from backpropagation. See more detail in BlockGrad or stop_gradient.

In addition, we can give a scale to the loss by setting grad_scale, so that the gradient of the loss will be rescaled in the backpropagation.

Note

This operator should be used as a Symbol instead of NDArray.

Defined in src/operator/make_loss.cc:L71

Parameters:
  • data (NDArray) – Input array.
  • grad_scale (float, optional, default=1) – Gradient scale as a supplement to unary and binary operators
  • valid_thresh (float, optional, default=0) – clip each element in the array to 0 when it is less than valid_thresh. This is used when normalization is set to 'valid'.
  • normalization ({'batch', 'null', 'valid'},optional, default='null') – If this is set to null, the output gradient will not be normalized. If this is set to batch, the output gradient will be divided by the batch size. If this is set to valid, the output gradient will be divided by the number of valid input elements.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Pad(data=None, mode=_Null, pad_width=_Null, constant_value=_Null, out=None, name=None, **kwargs)

Pads an input array with a constant or edge values of the array.

Note

Pad is deprecated. Use pad instead.

Note

Current implementation only supports 4D and 5D input arrays with padding applied only on axes 1, 2 and 3. Expects axes 4 and 5 in pad_width to be zero.

This operation pads an input array with either a constant_value or edge values along each axis of the input array. The amount of padding is specified by pad_width.

pad_width is a tuple of integer padding widths for each axis of the format (before_1, after_1, ... , before_N, after_N). The pad_width should be of length 2*N where N is the number of dimensions of the array.

For dimension N of the input array, before_N and after_N indicates how many values to add before and after the elements of the array along dimension N. The widths of the higher two dimensions before_1, after_1, before_2, after_2 must be 0.

Example:

x = [[[[  1.   2.   3.]
       [  4.   5.   6.]]

      [[  7.   8.   9.]
       [ 10.  11.  12.]]]


     [[[ 11.  12.  13.]
       [ 14.  15.  16.]]

      [[ 17.  18.  19.]
       [ 20.  21.  22.]]]]

pad(x,mode="edge", pad_width=(0,0,0,0,1,1,1,1)) =

      [[[[  1.   1.   2.   3.   3.]
         [  1.   1.   2.   3.   3.]
         [  4.   4.   5.   6.   6.]
         [  4.   4.   5.   6.   6.]]

        [[  7.   7.   8.   9.   9.]
         [  7.   7.   8.   9.   9.]
         [ 10.  10.  11.  12.  12.]
         [ 10.  10.  11.  12.  12.]]]


       [[[ 11.  11.  12.  13.  13.]
         [ 11.  11.  12.  13.  13.]
         [ 14.  14.  15.  16.  16.]
         [ 14.  14.  15.  16.  16.]]

        [[ 17.  17.  18.  19.  19.]
         [ 17.  17.  18.  19.  19.]
         [ 20.  20.  21.  22.  22.]
         [ 20.  20.  21.  22.  22.]]]]

pad(x, mode="constant", constant_value=0, pad_width=(0,0,0,0,1,1,1,1)) =

      [[[[  0.   0.   0.   0.   0.]
         [  0.   1.   2.   3.   0.]
         [  0.   4.   5.   6.   0.]
         [  0.   0.   0.   0.   0.]]

        [[  0.   0.   0.   0.   0.]
         [  0.   7.   8.   9.   0.]
         [  0.  10.  11.  12.   0.]
         [  0.   0.   0.   0.   0.]]]


       [[[  0.   0.   0.   0.   0.]
         [  0.  11.  12.  13.   0.]
         [  0.  14.  15.  16.   0.]
         [  0.   0.   0.   0.   0.]]

        [[  0.   0.   0.   0.   0.]
         [  0.  17.  18.  19.   0.]
         [  0.  20.  21.  22.   0.]
         [  0.   0.   0.   0.   0.]]]]

Defined in src/operator/pad.cc:L766

Parameters:
  • data (NDArray) – An n-dimensional input array.
  • mode ({'constant', 'edge', 'reflect'}, required) – Padding type to use. “constant” pads with constant_value “edge” pads using the edge values of the input array “reflect” pads by reflecting values with respect to the edges.
  • pad_width (Shape(tuple), required) – Widths of the padding regions applied to the edges of each axis. It is a tuple of integer padding widths for each axis of the format (before_1, after_1, ... , before_N, after_N). It should be of length 2*N where N is the number of dimensions of the array.This is equivalent to pad_width in numpy.pad, but flattened.
  • constant_value (double, optional, default=0) – The value used for padding when mode is “constant”.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Pooling(data=None, kernel=_Null, pool_type=_Null, global_pool=_Null, cudnn_off=_Null, pooling_convention=_Null, stride=_Null, pad=_Null, p_value=_Null, count_include_pad=_Null, layout=_Null, out=None, name=None, **kwargs)

Performs pooling on the input.

The shapes for 1-D pooling are

  • data and out: (batch_size, channel, width) (NCW layout) or (batch_size, width, channel) (NWC layout),

The shapes for 2-D pooling are

  • data and out: (batch_size, channel, height, width) (NCHW layout) or (batch_size, height, width, channel) (NHWC layout),

    out_height = f(height, kernel[0], pad[0], stride[0]) out_width = f(width, kernel[1], pad[1], stride[1])

The definition of f depends on pooling_convention, which has two options:

  • valid (default):

    f(x, k, p, s) = floor((x+2*p-k)/s)+1
    
  • full, which is compatible with Caffe:

    f(x, k, p, s) = ceil((x+2*p-k)/s)+1
    

But global_pool is set to be true, then do a global pooling, namely reset kernel=(height, width).

Three pooling options are supported by pool_type:

  • avg: average pooling
  • max: max pooling
  • sum: sum pooling
  • lp: Lp pooling

For 3-D pooling, an additional depth dimension is added before height. Namely the input data and output will have shape (batch_size, channel, depth, height, width) (NCDHW layout) or (batch_size, depth, height, width, channel) (NDHWC layout).

Notes on Lp pooling:

Lp pooling was first introduced by this paper: https://arxiv.org/pdf/1204.3968.pdf. L-1 pooling is simply sum pooling, while L-inf pooling is simply max pooling. We can see that Lp pooling stands between those two, in practice the most common value for p is 2.

For each window X, the mathematical expression for Lp pooling is:

\(f(X) = \sqrt[p]{\sum_{x}^{X} x^p}\)

Defined in src/operator/nn/pooling.cc:L416

Parameters:
  • data (NDArray) – Input data to the pooling operator.
  • kernel (Shape(tuple), optional, default=[]) – Pooling kernel size: (y, x) or (d, y, x)
  • pool_type ({'avg', 'lp', 'max', 'sum'},optional, default='max') – Pooling type to be applied.
  • global_pool (boolean, optional, default=0) – Ignore kernel size, do global pooling based on current input feature map.
  • cudnn_off (boolean, optional, default=0) – Turn off cudnn pooling and use MXNet pooling operator.
  • pooling_convention ({'full', 'same', 'valid'},optional, default='valid') – Pooling convention to be applied.
  • stride (Shape(tuple), optional, default=[]) – Stride: for pooling (y, x) or (d, y, x). Defaults to 1 for each dimension.
  • pad (Shape(tuple), optional, default=[]) – Pad for pooling: (y, x) or (d, y, x). Defaults to no padding.
  • p_value (int or None, optional, default='None') – Value of p for Lp pooling, can be 1 or 2, required for Lp Pooling.
  • count_include_pad (boolean or None, optional, default=None) – Only used for AvgPool, specify whether to count padding elements for averagecalculation. For example, with a 5*5 kernel on a 3*3 corner of a image,the sum of the 9 valid elements will be divided by 25 if this is set to true,or it will be divided by 9 if this is set to false. Defaults to true.
  • layout ({None, 'NCDHW', 'NCHW', 'NCW', 'NDHWC', 'NHWC', 'NWC'},optional, default='None') – Set layout for input and output. Empty for default layout: NCW for 1d, NCHW for 2d and NCDHW for 3d.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Pooling_v1(data=None, kernel=_Null, pool_type=_Null, global_pool=_Null, pooling_convention=_Null, stride=_Null, pad=_Null, out=None, name=None, **kwargs)

This operator is DEPRECATED. Perform pooling on the input.

The shapes for 2-D pooling is

  • data: (batch_size, channel, height, width)

  • out: (batch_size, num_filter, out_height, out_width), with:

    out_height = f(height, kernel[0], pad[0], stride[0])
    out_width = f(width, kernel[1], pad[1], stride[1])
    

The definition of f depends on pooling_convention, which has two options:

  • valid (default):

    f(x, k, p, s) = floor((x+2*p-k)/s)+1
    
  • full, which is compatible with Caffe:

    f(x, k, p, s) = ceil((x+2*p-k)/s)+1
    

But global_pool is set to be true, then do a global pooling, namely reset kernel=(height, width).

Three pooling options are supported by pool_type:

  • avg: average pooling
  • max: max pooling
  • sum: sum pooling

1-D pooling is special case of 2-D pooling with weight=1 and kernel[1]=1.

For 3-D pooling, an additional depth dimension is added before height. Namely the input data will have shape (batch_size, channel, depth, height, width).

Defined in src/operator/pooling_v1.cc:L104

Parameters:
  • data (NDArray) – Input data to the pooling operator.
  • kernel (Shape(tuple), optional, default=[]) – pooling kernel size: (y, x) or (d, y, x)
  • pool_type ({'avg', 'max', 'sum'},optional, default='max') – Pooling type to be applied.
  • global_pool (boolean, optional, default=0) – Ignore kernel size, do global pooling based on current input feature map.
  • pooling_convention ({'full', 'valid'},optional, default='valid') – Pooling convention to be applied.
  • stride (Shape(tuple), optional, default=[]) – stride: for pooling (y, x) or (d, y, x)
  • pad (Shape(tuple), optional, default=[]) – pad for pooling: (y, x) or (d, y, x)
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.RNN(data=None, parameters=None, state=None, state_cell=None, sequence_length=None, state_size=_Null, num_layers=_Null, bidirectional=_Null, mode=_Null, p=_Null, state_outputs=_Null, projection_size=_Null, lstm_state_clip_min=_Null, lstm_state_clip_max=_Null, lstm_state_clip_nan=_Null, use_sequence_length=_Null, out=None, name=None, **kwargs)

Applies recurrent layers to input data. Currently, vanilla RNN, LSTM and GRU are implemented, with both multi-layer and bidirectional support.

When the input data is of type float32 and the environment variables MXNET_CUDA_ALLOW_TENSOR_CORE and MXNET_CUDA_TENSOR_OP_MATH_ALLOW_CONVERSION are set to 1, this operator will try to use pseudo-float16 precision (float32 math with float16 I/O) precision in order to use Tensor Cores on suitable NVIDIA GPUs. This can sometimes give significant speedups.

Vanilla RNN

Applies a single-gate recurrent layer to input X. Two kinds of activation function are supported: ReLU and Tanh.

With ReLU activation function:

\[h_t = relu(W_{ih} * x_t + b_{ih} + W_{hh} * h_{(t-1)} + b_{hh})\]

With Tanh activtion function:

\[h_t = \tanh(W_{ih} * x_t + b_{ih} + W_{hh} * h_{(t-1)} + b_{hh})\]

Reference paper: Finding structure in time - Elman, 1988. https://crl.ucsd.edu/~elman/Papers/fsit.pdf

LSTM

Long Short-Term Memory - Hochreiter, 1997. http://www.bioinf.jku.at/publications/older/2604.pdf

\[\begin{split}\begin{array}{ll} i_t = \mathrm{sigmoid}(W_{ii} x_t + b_{ii} + W_{hi} h_{(t-1)} + b_{hi}) \\ f_t = \mathrm{sigmoid}(W_{if} x_t + b_{if} + W_{hf} h_{(t-1)} + b_{hf}) \\ g_t = \tanh(W_{ig} x_t + b_{ig} + W_{hc} h_{(t-1)} + b_{hg}) \\ o_t = \mathrm{sigmoid}(W_{io} x_t + b_{io} + W_{ho} h_{(t-1)} + b_{ho}) \\ c_t = f_t * c_{(t-1)} + i_t * g_t \\ h_t = o_t * \tanh(c_t) \end{array}\end{split}\]

GRU

Gated Recurrent Unit - Cho et al. 2014. http://arxiv.org/abs/1406.1078

The definition of GRU here is slightly different from paper but compatible with CUDNN.

\[\begin{split}\begin{array}{ll} r_t = \mathrm{sigmoid}(W_{ir} x_t + b_{ir} + W_{hr} h_{(t-1)} + b_{hr}) \\ z_t = \mathrm{sigmoid}(W_{iz} x_t + b_{iz} + W_{hz} h_{(t-1)} + b_{hz}) \\ n_t = \tanh(W_{in} x_t + b_{in} + r_t * (W_{hn} h_{(t-1)}+ b_{hn})) \\ h_t = (1 - z_t) * n_t + z_t * h_{(t-1)} \\ \end{array}\end{split}\]

Defined in src/operator/rnn.cc:L690

Parameters:
  • data (NDArray) – Input data to RNN
  • parameters (NDArray) – Vector of all RNN trainable parameters concatenated
  • state (NDArray) – initial hidden state of the RNN
  • state_cell (NDArray) – initial cell state for LSTM networks (only for LSTM)
  • sequence_length (NDArray) – Vector of valid sequence lengths for each element in batch. (Only used if use_sequence_length kwarg is True)
  • state_size (int (non-negative), required) – size of the state for each layer
  • num_layers (int (non-negative), required) – number of stacked layers
  • bidirectional (boolean, optional, default=0) – whether to use bidirectional recurrent layers
  • mode ({'gru', 'lstm', 'rnn_relu', 'rnn_tanh'}, required) – the type of RNN to compute
  • p (float, optional, default=0) – drop rate of the dropout on the outputs of each RNN layer, except the last layer.
  • state_outputs (boolean, optional, default=0) – Whether to have the states as symbol outputs.
  • projection_size (int or None, optional, default='None') – size of project size
  • lstm_state_clip_min (double or None, optional, default=None) – Minimum clip value of LSTM states. This option must be used together with lstm_state_clip_max.
  • lstm_state_clip_max (double or None, optional, default=None) – Maximum clip value of LSTM states. This option must be used together with lstm_state_clip_min.
  • lstm_state_clip_nan (boolean, optional, default=0) – Whether to stop NaN from propagating in state by clipping it to min/max. If clipping range is not specified, this option is ignored.
  • use_sequence_length (boolean, optional, default=0) – If set to true, this layer takes in an extra input parameter sequence_length to specify variable length sequence
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ROIPooling(data=None, rois=None, pooled_size=_Null, spatial_scale=_Null, out=None, name=None, **kwargs)

Performs region of interest(ROI) pooling on the input array.

ROI pooling is a variant of a max pooling layer, in which the output size is fixed and region of interest is a parameter. Its purpose is to perform max pooling on the inputs of non-uniform sizes to obtain fixed-size feature maps. ROI pooling is a neural-net layer mostly used in training a Fast R-CNN network for object detection.

This operator takes a 4D feature map as an input array and region proposals as rois, then it pools over sub-regions of input and produces a fixed-sized output array regardless of the ROI size.

To crop the feature map accordingly, you can resize the bounding box coordinates by changing the parameters rois and spatial_scale.

The cropped feature maps are pooled by standard max pooling operation to a fixed size output indicated by a pooled_size parameter. batch_size will change to the number of region bounding boxes after ROIPooling.

The size of each region of interest doesn’t have to be perfectly divisible by the number of pooling sections(pooled_size).

Example:

x = [[[[  0.,   1.,   2.,   3.,   4.,   5.],
       [  6.,   7.,   8.,   9.,  10.,  11.],
       [ 12.,  13.,  14.,  15.,  16.,  17.],
       [ 18.,  19.,  20.,  21.,  22.,  23.],
       [ 24.,  25.,  26.,  27.,  28.,  29.],
       [ 30.,  31.,  32.,  33.,  34.,  35.],
       [ 36.,  37.,  38.,  39.,  40.,  41.],
       [ 42.,  43.,  44.,  45.,  46.,  47.]]]]

// region of interest i.e. bounding box coordinates.
y = [[0,0,0,4,4]]

// returns array of shape (2,2) according to the given roi with max pooling.
ROIPooling(x, y, (2,2), 1.0) = [[[[ 14.,  16.],
                                  [ 26.,  28.]]]]

// region of interest is changed due to the change in `spacial_scale` parameter.
ROIPooling(x, y, (2,2), 0.7) = [[[[  7.,   9.],
                                  [ 19.,  21.]]]]

Defined in src/operator/roi_pooling.cc:L295

Parameters:
  • data (NDArray) – The input array to the pooling operator, a 4D Feature maps
  • rois (NDArray) – Bounding box coordinates, a 2D array of [[batch_index, x1, y1, x2, y2]], where (x1, y1) and (x2, y2) are top left and bottom right corners of designated region of interest. batch_index indicates the index of corresponding image in the input array
  • pooled_size (Shape(tuple), required) – ROI pooling output shape (h,w)
  • spatial_scale (float, required) – Ratio of input feature map height (or w) to raw image height (or w). Equals the reciprocal of total stride in convolutional layers
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Reshape(data=None, shape=_Null, reverse=_Null, target_shape=_Null, keep_highest=_Null, out=None, name=None, **kwargs)

Reshapes the input array.

Note

Reshape is deprecated, use reshape

Given an array and a shape, this function returns a copy of the array in the new shape. The shape is a tuple of integers such as (2,3,4). The size of the new shape should be same as the size of the input array.

Example:

reshape([1,2,3,4], shape=(2,2)) = [[1,2], [3,4]]

Some dimensions of the shape can take special values from the set {0, -1, -2, -3, -4}. The significance of each is explained below:

  • 0 copy this dimension from the input to the output shape.

    Example:

    - input shape = (2,3,4), shape = (4,0,2), output shape = (4,3,2)
    - input shape = (2,3,4), shape = (2,0,0), output shape = (2,3,4)
    
  • -1 infers the dimension of the output shape by using the remainder of the input dimensions keeping the size of the new array same as that of the input array. At most one dimension of shape can be -1.

    Example:

    - input shape = (2,3,4), shape = (6,1,-1), output shape = (6,1,4)
    - input shape = (2,3,4), shape = (3,-1,8), output shape = (3,1,8)
    - input shape = (2,3,4), shape=(-1,), output shape = (24,)
    
  • -2 copy all/remainder of the input dimensions to the output shape.

    Example:

    - input shape = (2,3,4), shape = (-2,), output shape = (2,3,4)
    - input shape = (2,3,4), shape = (2,-2), output shape = (2,3,4)
    - input shape = (2,3,4), shape = (-2,1,1), output shape = (2,3,4,1,1)
    
  • -3 use the product of two consecutive dimensions of the input shape as the output dimension.

    Example:

    - input shape = (2,3,4), shape = (-3,4), output shape = (6,4)
    - input shape = (2,3,4,5), shape = (-3,-3), output shape = (6,20)
    - input shape = (2,3,4), shape = (0,-3), output shape = (2,12)
    - input shape = (2,3,4), shape = (-3,-2), output shape = (6,4)
    
  • -4 split one dimension of the input into two dimensions passed subsequent to -4 in shape (can contain -1).

    Example:

    - input shape = (2,3,4), shape = (-4,1,2,-2), output shape =(1,2,3,4)
    - input shape = (2,3,4), shape = (2,-4,-1,3,-2), output shape = (2,1,3,4)
    

If the argument reverse is set to 1, then the special values are inferred from right to left.

Example:

- without reverse=1, for input shape = (10,5,4), shape = (-1,0), output shape would be (40,5)
- with reverse=1, output shape will be (50,4).

Defined in src/operator/tensor/matrix_op.cc:L202

Parameters:
  • data (NDArray) – Input data to reshape.
  • shape (Shape(tuple), optional, default=[]) – The target shape
  • reverse (boolean, optional, default=0) – If true then the special values are inferred from right to left
  • target_shape (Shape(tuple), optional, default=[]) – (Deprecated! Use shape instead.) Target new shape. One and only one dim can be 0, in which case it will be inferred from the rest of dims
  • keep_highest (boolean, optional, default=0) – (Deprecated! Use shape instead.) Whether keep the highest dim unchanged.If set to true, then the first dim in target_shape is ignored,and always fixed as input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

Examples

Reshapes the input array into a new shape.

>>> x = mx.nd.array([1, 2, 3, 4])
>>> y = mx.nd.reshape(x, shape=(2, 2))
>>> x.shape
(4L,)
>>> y.shape
(2L, 2L)
>>> y.asnumpy()
array([[ 1.,  2.],
   [ 3.,  4.]], dtype=float32)

You can use 0 to copy a particular dimension from the input to the output shape and ‘-1’ to infer the dimensions of the output.

>>> x = mx.nd.ones((2, 3, 4))
>>> x.shape
(2L, 3L, 4L)
>>> y = mx.nd.reshape(x, shape=(4, 0, -1))
>>> y.shape
(4L, 3L, 2L)
mxnet.ndarray.SVMOutput(data=None, label=None, margin=_Null, regularization_coefficient=_Null, use_linear=_Null, out=None, name=None, **kwargs)

Computes support vector machine based transformation of the input.

This tutorial demonstrates using SVM as output layer for classification instead of softmax: https://github.com/dmlc/mxnet/tree/master/example/svm_mnist.

Parameters:
  • data (NDArray) – Input data for SVM transformation.
  • label (NDArray) – Class label for the input data.
  • margin (float, optional, default=1) – The loss function penalizes outputs that lie outside this margin. Default margin is 1.
  • regularization_coefficient (float, optional, default=1) – Regularization parameter for the SVM. This balances the tradeoff between coefficient size and error.
  • use_linear (boolean, optional, default=0) – Whether to use L1-SVM objective. L2-SVM objective is used by default.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SequenceLast(data=None, sequence_length=None, use_sequence_length=_Null, axis=_Null, out=None, name=None, **kwargs)

Takes the last element of a sequence.

This function takes an n-dimensional input array of the form [max_sequence_length, batch_size, other_feature_dims] and returns a (n-1)-dimensional array of the form [batch_size, other_feature_dims].

Parameter sequence_length is used to handle variable-length sequences. sequence_length should be an input array of positive ints of dimension [batch_size]. To use this parameter, set use_sequence_length to True, otherwise each example in the batch is assumed to have the max sequence length.

Note

Alternatively, you can also use take operator.

Example:

x = [[[  1.,   2.,   3.],
      [  4.,   5.,   6.],
      [  7.,   8.,   9.]],

     [[ 10.,   11.,   12.],
      [ 13.,   14.,   15.],
      [ 16.,   17.,   18.]],

     [[  19.,   20.,   21.],
      [  22.,   23.,   24.],
      [  25.,   26.,   27.]]]

// returns last sequence when sequence_length parameter is not used
SequenceLast(x) = [[  19.,   20.,   21.],
                   [  22.,   23.,   24.],
                   [  25.,   26.,   27.]]

// sequence_length is used
SequenceLast(x, sequence_length=[1,1,1], use_sequence_length=True) =
         [[  1.,   2.,   3.],
          [  4.,   5.,   6.],
          [  7.,   8.,   9.]]

// sequence_length is used
SequenceLast(x, sequence_length=[1,2,3], use_sequence_length=True) =
         [[  1.,    2.,   3.],
          [  13.,  14.,  15.],
          [  25.,  26.,  27.]]

Defined in src/operator/sequence_last.cc:L100

Parameters:
  • data (NDArray) – n-dimensional input array of the form [max_sequence_length, batch_size, other_feature_dims] where n>2
  • sequence_length (NDArray) – vector of sequence lengths of the form [batch_size]
  • use_sequence_length (boolean, optional, default=0) – If set to true, this layer takes in an extra input parameter sequence_length to specify variable length sequence
  • axis (int, optional, default='0') – The sequence axis. Only values of 0 and 1 are currently supported.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SequenceMask(data=None, sequence_length=None, use_sequence_length=_Null, value=_Null, axis=_Null, out=None, name=None, **kwargs)

Sets all elements outside the sequence to a constant value.

This function takes an n-dimensional input array of the form [max_sequence_length, batch_size, other_feature_dims] and returns an array of the same shape.

Parameter sequence_length is used to handle variable-length sequences. sequence_length should be an input array of positive ints of dimension [batch_size]. To use this parameter, set use_sequence_length to True, otherwise each example in the batch is assumed to have the max sequence length and this operator works as the identity operator.

Example:

x = [[[  1.,   2.,   3.],
      [  4.,   5.,   6.]],

     [[  7.,   8.,   9.],
      [ 10.,  11.,  12.]],

     [[ 13.,  14.,   15.],
      [ 16.,  17.,   18.]]]

// Batch 1
B1 = [[  1.,   2.,   3.],
      [  7.,   8.,   9.],
      [ 13.,  14.,  15.]]

// Batch 2
B2 = [[  4.,   5.,   6.],
      [ 10.,  11.,  12.],
      [ 16.,  17.,  18.]]

// works as identity operator when sequence_length parameter is not used
SequenceMask(x) = [[[  1.,   2.,   3.],
                    [  4.,   5.,   6.]],

                   [[  7.,   8.,   9.],
                    [ 10.,  11.,  12.]],

                   [[ 13.,  14.,   15.],
                    [ 16.,  17.,   18.]]]

// sequence_length [1,1] means 1 of each batch will be kept
// and other rows are masked with default mask value = 0
SequenceMask(x, sequence_length=[1,1], use_sequence_length=True) =
             [[[  1.,   2.,   3.],
               [  4.,   5.,   6.]],

              [[  0.,   0.,   0.],
               [  0.,   0.,   0.]],

              [[  0.,   0.,   0.],
               [  0.,   0.,   0.]]]

// sequence_length [2,3] means 2 of batch B1 and 3 of batch B2 will be kept
// and other rows are masked with value = 1
SequenceMask(x, sequence_length=[2,3], use_sequence_length=True, value=1) =
             [[[  1.,   2.,   3.],
               [  4.,   5.,   6.]],

              [[  7.,   8.,   9.],
               [  10.,  11.,  12.]],

              [[   1.,   1.,   1.],
               [  16.,  17.,  18.]]]

Defined in src/operator/sequence_mask.cc:L186

Parameters:
  • data (NDArray) – n-dimensional input array of the form [max_sequence_length, batch_size, other_feature_dims] where n>2
  • sequence_length (NDArray) – vector of sequence lengths of the form [batch_size]
  • use_sequence_length (boolean, optional, default=0) – If set to true, this layer takes in an extra input parameter sequence_length to specify variable length sequence
  • value (float, optional, default=0) – The value to be used as a mask.
  • axis (int, optional, default='0') – The sequence axis. Only values of 0 and 1 are currently supported.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SequenceReverse(data=None, sequence_length=None, use_sequence_length=_Null, axis=_Null, out=None, name=None, **kwargs)

Reverses the elements of each sequence.

This function takes an n-dimensional input array of the form [max_sequence_length, batch_size, other_feature_dims] and returns an array of the same shape.

Parameter sequence_length is used to handle variable-length sequences. sequence_length should be an input array of positive ints of dimension [batch_size]. To use this parameter, set use_sequence_length to True, otherwise each example in the batch is assumed to have the max sequence length.

Example:

x = [[[  1.,   2.,   3.],
      [  4.,   5.,   6.]],

     [[  7.,   8.,   9.],
      [ 10.,  11.,  12.]],

     [[ 13.,  14.,   15.],
      [ 16.,  17.,   18.]]]

// Batch 1
B1 = [[  1.,   2.,   3.],
      [  7.,   8.,   9.],
      [ 13.,  14.,  15.]]

// Batch 2
B2 = [[  4.,   5.,   6.],
      [ 10.,  11.,  12.],
      [ 16.,  17.,  18.]]

// returns reverse sequence when sequence_length parameter is not used
SequenceReverse(x) = [[[ 13.,  14.,   15.],
                       [ 16.,  17.,   18.]],

                      [[  7.,   8.,   9.],
                       [ 10.,  11.,  12.]],

                      [[  1.,   2.,   3.],
                       [  4.,   5.,   6.]]]

// sequence_length [2,2] means 2 rows of
// both batch B1 and B2 will be reversed.
SequenceReverse(x, sequence_length=[2,2], use_sequence_length=True) =
                  [[[  7.,   8.,   9.],
                    [ 10.,  11.,  12.]],

                   [[  1.,   2.,   3.],
                    [  4.,   5.,   6.]],

                   [[ 13.,  14.,   15.],
                    [ 16.,  17.,   18.]]]

// sequence_length [2,3] means 2 of batch B2 and 3 of batch B3
// will be reversed.
SequenceReverse(x, sequence_length=[2,3], use_sequence_length=True) =
                 [[[  7.,   8.,   9.],
                   [ 16.,  17.,  18.]],

                  [[  1.,   2.,   3.],
                   [ 10.,  11.,  12.]],

                  [[ 13.,  14,   15.],
                   [  4.,   5.,   6.]]]

Defined in src/operator/sequence_reverse.cc:L122

Parameters:
  • data (NDArray) – n-dimensional input array of the form [max_sequence_length, batch_size, other dims] where n>2
  • sequence_length (NDArray) – vector of sequence lengths of the form [batch_size]
  • use_sequence_length (boolean, optional, default=0) – If set to true, this layer takes in an extra input parameter sequence_length to specify variable length sequence
  • axis (int, optional, default='0') – The sequence axis. Only 0 is currently supported.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SliceChannel(data=None, num_outputs=_Null, axis=_Null, squeeze_axis=_Null, out=None, name=None, **kwargs)

Splits an array along a particular axis into multiple sub-arrays.

Note

SliceChannel is deprecated. Use split instead.

Note that num_outputs should evenly divide the length of the axis along which to split the array.

Example:

x  = [[[ 1.]
       [ 2.]]
      [[ 3.]
       [ 4.]]
      [[ 5.]
       [ 6.]]]
x.shape = (3, 2, 1)

y = split(x, axis=1, num_outputs=2) // a list of 2 arrays with shape (3, 1, 1)
y = [[[ 1.]]
     [[ 3.]]
     [[ 5.]]]

    [[[ 2.]]
     [[ 4.]]
     [[ 6.]]]

y[0].shape = (3, 1, 1)

z = split(x, axis=0, num_outputs=3) // a list of 3 arrays with shape (1, 2, 1)
z = [[[ 1.]
      [ 2.]]]

    [[[ 3.]
      [ 4.]]]

    [[[ 5.]
      [ 6.]]]

z[0].shape = (1, 2, 1)

squeeze_axis=1 removes the axis with length 1 from the shapes of the output arrays. Note that setting squeeze_axis to 1 removes axis with length 1 only along the axis which it is split. Also squeeze_axis can be set to true only if input.shape[axis] == num_outputs.

Example:

z = split(x, axis=0, num_outputs=3, squeeze_axis=1) // a list of 3 arrays with shape (2, 1)
z = [[ 1.]
     [ 2.]]

    [[ 3.]
     [ 4.]]

    [[ 5.]
     [ 6.]]
z[0].shape = (2 ,1 )

Defined in src/operator/slice_channel.cc:L107

Parameters:
  • data (NDArray) – The input
  • num_outputs (int, required) – Number of splits. Note that this should evenly divide the length of the axis.
  • axis (int, optional, default='1') – Axis along which to split.
  • squeeze_axis (boolean, optional, default=0) – If true, Removes the axis with length 1 from the shapes of the output arrays. Note that setting squeeze_axis to true removes axis with length 1 only along the axis which it is split. Also squeeze_axis can be set to true only if input.shape[axis] == num_outputs.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.Softmax(data=None, label=None, grad_scale=_Null, ignore_label=_Null, multi_output=_Null, use_ignore=_Null, preserve_shape=_Null, normalization=_Null, out_grad=_Null, smooth_alpha=_Null, out=None, name=None, **kwargs)

Computes the gradient of cross entropy loss with respect to softmax output.

  • This operator computes the gradient in two steps. The cross entropy loss does not actually need to be computed.

    • Applies softmax function on the input array.
    • Computes and returns the gradient of cross entropy loss w.r.t. the softmax output.
  • The softmax function, cross entropy loss and gradient is given by:

    • Softmax Function:

      \[\text{softmax}(x)_i = \frac{exp(x_i)}{\sum_j exp(x_j)}\]
    • Cross Entropy Function:

      \[\text{CE(label, output)} = - \sum_i \text{label}_i \log(\text{output}_i)\]
    • The gradient of cross entropy loss w.r.t softmax output:

      \[\text{gradient} = \text{output} - \text{label}\]
  • During forward propagation, the softmax function is computed for each instance in the input array.

    For general N-D input arrays with shape \((d_1, d_2, ..., d_n)\). The size is \(s=d_1 \cdot d_2 \cdot \cdot \cdot d_n\). We can use the parameters preserve_shape and multi_output to specify the way to compute softmax:

    • By default, preserve_shape is false. This operator will reshape the input array into a 2-D array with shape \((d_1, \frac{s}{d_1})\) and then compute the softmax function for each row in the reshaped array, and afterwards reshape it back to the original shape \((d_1, d_2, ..., d_n)\).
    • If preserve_shape is true, the softmax function will be computed along the last axis (axis = -1).
    • If multi_output is true, the softmax function will be computed along the second axis (axis = 1).
  • During backward propagation, the gradient of cross-entropy loss w.r.t softmax output array is computed. The provided label can be a one-hot label array or a probability label array.

    • If the parameter use_ignore is true, ignore_label can specify input instances with a particular label to be ignored during backward propagation. This has no effect when softmax `output` has same shape as `label`.

      Example:

      data = [[1,2,3,4],[2,2,2,2],[3,3,3,3],[4,4,4,4]]
      label = [1,0,2,3]
      ignore_label = 1
      SoftmaxOutput(data=data, label = label,\
                    multi_output=true, use_ignore=true,\
                    ignore_label=ignore_label)
      ## forward softmax output
      [[ 0.0320586   0.08714432  0.23688284  0.64391428]
       [ 0.25        0.25        0.25        0.25      ]
       [ 0.25        0.25        0.25        0.25      ]
       [ 0.25        0.25        0.25        0.25      ]]
      ## backward gradient output
      [[ 0.    0.    0.    0.  ]
       [-0.75  0.25  0.25  0.25]
       [ 0.25  0.25 -0.75  0.25]
       [ 0.25  0.25  0.25 -0.75]]
      ## notice that the first row is all 0 because label[0] is 1, which is equal to ignore_label.
      
    • The parameter grad_scale can be used to rescale the gradient, which is often used to give each loss function different weights.

    • This operator also supports various ways to normalize the gradient by normalization, The normalization is applied if softmax output has different shape than the labels. The normalization mode can be set to the followings:

      • 'null': do nothing.
      • 'batch': divide the gradient by the batch size.
      • 'valid': divide the gradient by the number of instances which are not ignored.

Defined in src/operator/softmax_output.cc:L230

Parameters:
  • data (NDArray) – Input array.
  • label (NDArray) – Ground truth label.
  • grad_scale (float, optional, default=1) – Scales the gradient by a float factor.
  • ignore_label (float, optional, default=-1) – The instances whose labels == ignore_label will be ignored during backward, if use_ignore is set to true).
  • multi_output (boolean, optional, default=0) – If set to true, the softmax function will be computed along axis 1. This is applied when the shape of input array differs from the shape of label array.
  • use_ignore (boolean, optional, default=0) – If set to true, the ignore_label value will not contribute to the backward gradient.
  • preserve_shape (boolean, optional, default=0) – If set to true, the softmax function will be computed along the last axis (-1).
  • normalization ({'batch', 'null', 'valid'},optional, default='null') – Normalizes the gradient.
  • out_grad (boolean, optional, default=0) – Multiplies gradient with output gradient element-wise.
  • smooth_alpha (float, optional, default=0) – Constant for computing a label smoothed version of cross-entropyfor the backwards pass. This constant gets subtracted from theone-hot encoding of the gold label and distributed uniformly toall other labels.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SoftmaxActivation(data=None, mode=_Null, out=None, name=None, **kwargs)

Applies softmax activation to input. This is intended for internal layers.

Note

This operator has been deprecated, please use softmax.

If mode = instance, this operator will compute a softmax for each instance in the batch. This is the default mode.

If mode = channel, this operator will compute a k-class softmax at each position of each instance, where k = num_channel. This mode can only be used when the input array has at least 3 dimensions. This can be used for fully convolutional network, image segmentation, etc.

Example:

>>> input_array = mx.nd.array([[3., 0.5, -0.5, 2., 7.],
>>>                            [2., -.4, 7.,   3., 0.2]])
>>> softmax_act = mx.nd.SoftmaxActivation(input_array)
>>> print softmax_act.asnumpy()
[[  1.78322066e-02   1.46375655e-03   5.38485940e-04   6.56010211e-03   9.73605454e-01]
 [  6.56221947e-03   5.95310994e-04   9.73919690e-01   1.78379621e-02   1.08472735e-03]]

Defined in src/operator/nn/softmax_activation.cc:L59

Parameters:
  • data (NDArray) – The input array.
  • mode ({'channel', 'instance'},optional, default='instance') – Specifies how to compute the softmax. If set to instance, it computes softmax for each instance. If set to channel, It computes cross channel softmax for each position of each instance.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SoftmaxOutput(data=None, label=None, grad_scale=_Null, ignore_label=_Null, multi_output=_Null, use_ignore=_Null, preserve_shape=_Null, normalization=_Null, out_grad=_Null, smooth_alpha=_Null, out=None, name=None, **kwargs)

Computes the gradient of cross entropy loss with respect to softmax output.

  • This operator computes the gradient in two steps. The cross entropy loss does not actually need to be computed.

    • Applies softmax function on the input array.
    • Computes and returns the gradient of cross entropy loss w.r.t. the softmax output.
  • The softmax function, cross entropy loss and gradient is given by:

    • Softmax Function:

      \[\text{softmax}(x)_i = \frac{exp(x_i)}{\sum_j exp(x_j)}\]
    • Cross Entropy Function:

      \[\text{CE(label, output)} = - \sum_i \text{label}_i \log(\text{output}_i)\]
    • The gradient of cross entropy loss w.r.t softmax output:

      \[\text{gradient} = \text{output} - \text{label}\]
  • During forward propagation, the softmax function is computed for each instance in the input array.

    For general N-D input arrays with shape \((d_1, d_2, ..., d_n)\). The size is \(s=d_1 \cdot d_2 \cdot \cdot \cdot d_n\). We can use the parameters preserve_shape and multi_output to specify the way to compute softmax:

    • By default, preserve_shape is false. This operator will reshape the input array into a 2-D array with shape \((d_1, \frac{s}{d_1})\) and then compute the softmax function for each row in the reshaped array, and afterwards reshape it back to the original shape \((d_1, d_2, ..., d_n)\).
    • If preserve_shape is true, the softmax function will be computed along the last axis (axis = -1).
    • If multi_output is true, the softmax function will be computed along the second axis (axis = 1).
  • During backward propagation, the gradient of cross-entropy loss w.r.t softmax output array is computed. The provided label can be a one-hot label array or a probability label array.

    • If the parameter use_ignore is true, ignore_label can specify input instances with a particular label to be ignored during backward propagation. This has no effect when softmax `output` has same shape as `label`.

      Example:

      data = [[1,2,3,4],[2,2,2,2],[3,3,3,3],[4,4,4,4]]
      label = [1,0,2,3]
      ignore_label = 1
      SoftmaxOutput(data=data, label = label,\
                    multi_output=true, use_ignore=true,\
                    ignore_label=ignore_label)
      ## forward softmax output
      [[ 0.0320586   0.08714432  0.23688284  0.64391428]
       [ 0.25        0.25        0.25        0.25      ]
       [ 0.25        0.25        0.25        0.25      ]
       [ 0.25        0.25        0.25        0.25      ]]
      ## backward gradient output
      [[ 0.    0.    0.    0.  ]
       [-0.75  0.25  0.25  0.25]
       [ 0.25  0.25 -0.75  0.25]
       [ 0.25  0.25  0.25 -0.75]]
      ## notice that the first row is all 0 because label[0] is 1, which is equal to ignore_label.
      
    • The parameter grad_scale can be used to rescale the gradient, which is often used to give each loss function different weights.

    • This operator also supports various ways to normalize the gradient by normalization, The normalization is applied if softmax output has different shape than the labels. The normalization mode can be set to the followings:

      • 'null': do nothing.
      • 'batch': divide the gradient by the batch size.
      • 'valid': divide the gradient by the number of instances which are not ignored.

Defined in src/operator/softmax_output.cc:L230

Parameters:
  • data (NDArray) – Input array.
  • label (NDArray) – Ground truth label.
  • grad_scale (float, optional, default=1) – Scales the gradient by a float factor.
  • ignore_label (float, optional, default=-1) – The instances whose labels == ignore_label will be ignored during backward, if use_ignore is set to true).
  • multi_output (boolean, optional, default=0) – If set to true, the softmax function will be computed along axis 1. This is applied when the shape of input array differs from the shape of label array.
  • use_ignore (boolean, optional, default=0) – If set to true, the ignore_label value will not contribute to the backward gradient.
  • preserve_shape (boolean, optional, default=0) – If set to true, the softmax function will be computed along the last axis (-1).
  • normalization ({'batch', 'null', 'valid'},optional, default='null') – Normalizes the gradient.
  • out_grad (boolean, optional, default=0) – Multiplies gradient with output gradient element-wise.
  • smooth_alpha (float, optional, default=0) – Constant for computing a label smoothed version of cross-entropyfor the backwards pass. This constant gets subtracted from theone-hot encoding of the gold label and distributed uniformly toall other labels.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SpatialTransformer(data=None, loc=None, target_shape=_Null, transform_type=_Null, sampler_type=_Null, cudnn_off=_Null, out=None, name=None, **kwargs)

Applies a spatial transformer to input feature map.

Parameters:
  • data (NDArray) – Input data to the SpatialTransformerOp.
  • loc (NDArray) – localisation net, the output dim should be 6 when transform_type is affine. You shold initialize the weight and bias with identity tranform.
  • target_shape (Shape(tuple), optional, default=[0,0]) – output shape(h, w) of spatial transformer: (y, x)
  • transform_type ({'affine'}, required) – transformation type
  • sampler_type ({'bilinear'}, required) – sampling type
  • cudnn_off (boolean or None, optional, default=None) – whether to turn cudnn off
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.SwapAxis(data=None, dim1=_Null, dim2=_Null, out=None, name=None, **kwargs)

Interchanges two axes of an array.

Examples:

 x = [[1, 2, 3]])
 swapaxes(x, 0, 1) = [[ 1],
                      [ 2],
                      [ 3]]

 x = [[[ 0, 1],
       [ 2, 3]],
      [[ 4, 5],
       [ 6, 7]]]  // (2,2,2) array

swapaxes(x, 0, 2) = [[[ 0, 4],
                      [ 2, 6]],
                     [[ 1, 5],
                      [ 3, 7]]]

Defined in src/operator/swapaxis.cc:L70

Parameters:
  • data (NDArray) – Input array.
  • dim1 (int (non-negative), optional, default=0) – the first axis to be swapped.
  • dim2 (int (non-negative), optional, default=0) – the second axis to be swapped.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.UpSampling(*data, **kwargs)

Upsamples the given input data.

Two algorithms (sample_type) are available for upsampling:

  • Nearest Neighbor
  • Bilinear

Nearest Neighbor Upsampling

Input data is expected to be NCHW.

Example:

x = [[[[1. 1. 1.]
       [1. 1. 1.]
       [1. 1. 1.]]]]

UpSampling(x, scale=2, sample_type='nearest') = [[[[1. 1. 1. 1. 1. 1.]
                                                   [1. 1. 1. 1. 1. 1.]
                                                   [1. 1. 1. 1. 1. 1.]
                                                   [1. 1. 1. 1. 1. 1.]
                                                   [1. 1. 1. 1. 1. 1.]
                                                   [1. 1. 1. 1. 1. 1.]]]]

Bilinear Upsampling

Uses deconvolution algorithm under the hood. You need provide both input data and the kernel.

Input data is expected to be NCHW.

num_filter is expected to be same as the number of channels.

Example:

x = [[[[1. 1. 1.]
       [1. 1. 1.]
       [1. 1. 1.]]]]

w = [[[[1. 1. 1. 1.]
       [1. 1. 1. 1.]
       [1. 1. 1. 1.]
       [1. 1. 1. 1.]]]]

UpSampling(x, w, scale=2, sample_type='bilinear', num_filter=1) = [[[[1. 2. 2. 2. 2. 1.]
                                                                     [2. 4. 4. 4. 4. 2.]
                                                                     [2. 4. 4. 4. 4. 2.]
                                                                     [2. 4. 4. 4. 4. 2.]
                                                                     [2. 4. 4. 4. 4. 2.]
                                                                     [1. 2. 2. 2. 2. 1.]]]]

Defined in src/operator/nn/upsampling.cc:L173

Parameters:
  • data (NDArray[]) – Array of tensors to upsample. For bilinear upsampling, there should be 2 inputs - 1 data and 1 weight.
  • scale (int, required) – Up sampling scale
  • num_filter (int, optional, default='0') – Input filter. Only used by bilinear sample_type.Since bilinear upsampling uses deconvolution, num_filters is set to the number of channels.
  • sample_type ({'bilinear', 'nearest'}, required) – upsampling method
  • multi_input_mode ({'concat', 'sum'},optional, default='concat') – How to handle multiple input. concat means concatenate upsampled images along the channel dimension. sum means add all images together, only available for nearest neighbor upsampling.
  • workspace (long (non-negative), optional, default=512) – Tmp workspace for deconvolution (MB)
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.abs(data=None, out=None, name=None, **kwargs)

Returns element-wise absolute value of the input.

Example:

abs([-2, 0, 3]) = [2, 0, 3]

The storage type of abs output depends upon the input storage type:

  • abs(default) = default
  • abs(row_sparse) = row_sparse
  • abs(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L708

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.adam_update(weight=None, grad=None, mean=None, var=None, lr=_Null, beta1=_Null, beta2=_Null, epsilon=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, lazy_update=_Null, out=None, name=None, **kwargs)

Update function for Adam optimizer. Adam is seen as a generalization of AdaGrad.

Adam update consists of the following steps, where g represents gradient and m, v are 1st and 2nd order moment estimates (mean and variance).

\[\begin{split}g_t = \nabla J(W_{t-1})\\ m_t = \beta_1 m_{t-1} + (1 - \beta_1) g_t\\ v_t = \beta_2 v_{t-1} + (1 - \beta_2) g_t^2\\ W_t = W_{t-1} - \alpha \frac{ m_t }{ \sqrt{ v_t } + \epsilon }\end{split}\]

It updates the weights using:

m = beta1*m + (1-beta1)*grad
v = beta2*v + (1-beta2)*(grad**2)
w += - learning_rate * m / (sqrt(v) + epsilon)

However, if grad’s storage type is row_sparse, lazy_update is True and the storage type of weight is the same as those of m and v, only the row slices whose indices appear in grad.indices are updated (for w, m and v):

for row in grad.indices:
    m[row] = beta1*m[row] + (1-beta1)*grad[row]
    v[row] = beta2*v[row] + (1-beta2)*(grad[row]**2)
    w[row] += - learning_rate * m[row] / (sqrt(v[row]) + epsilon)

Defined in src/operator/optimizer_op.cc:L686

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • mean (NDArray) – Moving mean
  • var (NDArray) – Moving variance
  • lr (float, required) – Learning rate
  • beta1 (float, optional, default=0.899999976) – The decay rate for the 1st moment estimates.
  • beta2 (float, optional, default=0.999000013) – The decay rate for the 2nd moment estimates.
  • epsilon (float, optional, default=9.99999994e-09) – A small constant for numerical stability.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • lazy_update (boolean, optional, default=1) – If true, lazy updates are applied if gradient’s stype is row_sparse and all of w, m and v have the same stype
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.add_n(*args, **kwargs)

Adds all input arguments element-wise.

\[add\_n(a_1, a_2, ..., a_n) = a_1 + a_2 + ... + a_n\]

add_n is potentially more efficient than calling add by n times.

The storage type of add_n output depends on storage types of inputs

  • add_n(row_sparse, row_sparse, ..) = row_sparse
  • add_n(default, csr, default) = default
  • add_n(any input combinations longer than 4 (>4) with at least one default type) = default
  • otherwise, add_n falls all inputs back to default storage and generates default storage

Defined in src/operator/tensor/elemwise_sum.cc:L155

Parameters:
  • args (NDArray[]) – Positional input arguments
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.all_finite(data=None, init_output=_Null, out=None, name=None, **kwargs)

Check if all the float numbers in the array are finite (used for AMP)

Defined in src/operator/contrib/all_finite.cc:L101

Parameters:
  • data (NDArray) – Array
  • init_output (boolean, optional, default=1) – Initialize output to 1.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.amp_cast(data=None, dtype=_Null, out=None, name=None, **kwargs)

Cast function between low precision float/FP32 used by AMP.

It casts only between low precision float/FP32 and does not do anything for other types.

Defined in src/operator/tensor/amp_cast.cc:L37

Parameters:
  • data (NDArray) – The input.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8'}, required) – Output data type.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.amp_multicast(*data, **kwargs)

Cast function used by AMP, that casts its inputs to the common widest type.

It casts only between low precision float/FP32 and does not do anything for other types.

Defined in src/operator/tensor/amp_cast.cc:L71

Parameters:
  • data (NDArray[]) – Weights
  • num_outputs (int, required) – Number of input/output pairs to be casted to the widest type.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.arccos(data=None, out=None, name=None, **kwargs)

Returns element-wise inverse cosine of the input array.

The input should be in range [-1, 1]. The output is in the closed interval \([0, \pi]\)

\[arccos([-1, -.707, 0, .707, 1]) = [\pi, 3\pi/4, \pi/2, \pi/4, 0]\]

The storage type of arccos output is always dense

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L179

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.arccosh(data=None, out=None, name=None, **kwargs)

Returns the element-wise inverse hyperbolic cosine of the input array, computed element-wise.

The storage type of arccosh output is always dense

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L320

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.arcsin(data=None, out=None, name=None, **kwargs)

Returns element-wise inverse sine of the input array.

The input should be in the range [-1, 1]. The output is in the closed interval of [\(-\pi/2\), \(\pi/2\)].

\[arcsin([-1, -.707, 0, .707, 1]) = [-\pi/2, -\pi/4, 0, \pi/4, \pi/2]\]

The storage type of arcsin output depends upon the input storage type:

  • arcsin(default) = default
  • arcsin(row_sparse) = row_sparse
  • arcsin(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L160

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.arcsinh(data=None, out=None, name=None, **kwargs)

Returns the element-wise inverse hyperbolic sine of the input array, computed element-wise.

The storage type of arcsinh output depends upon the input storage type:

  • arcsinh(default) = default
  • arcsinh(row_sparse) = row_sparse
  • arcsinh(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L306

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.arctan(data=None, out=None, name=None, **kwargs)

Returns element-wise inverse tangent of the input array.

The output is in the closed interval \([-\pi/2, \pi/2]\)

\[arctan([-1, 0, 1]) = [-\pi/4, 0, \pi/4]\]

The storage type of arctan output depends upon the input storage type:

  • arctan(default) = default
  • arctan(row_sparse) = row_sparse
  • arctan(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L200

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.arctanh(data=None, out=None, name=None, **kwargs)

Returns the element-wise inverse hyperbolic tangent of the input array, computed element-wise.

The storage type of arctanh output depends upon the input storage type:

  • arctanh(default) = default
  • arctanh(row_sparse) = row_sparse
  • arctanh(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L337

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.argmax(data=None, axis=_Null, keepdims=_Null, out=None, name=None, **kwargs)

Returns indices of the maximum values along an axis.

In the case of multiple occurrences of maximum values, the indices corresponding to the first occurrence are returned.

Examples:

x = [[ 0.,  1.,  2.],
     [ 3.,  4.,  5.]]

// argmax along axis 0
argmax(x, axis=0) = [ 1.,  1.,  1.]

// argmax along axis 1
argmax(x, axis=1) = [ 2.,  2.]

// argmax along axis 1 keeping same dims as an input array
argmax(x, axis=1, keepdims=True) = [[ 2.],
                                    [ 2.]]

Defined in src/operator/tensor/broadcast_reduce_op_index.cc:L52

Parameters:
  • data (NDArray) – The input
  • axis (int or None, optional, default='None') – The axis along which to perform the reduction. Negative values means indexing from right to left. Requires axis to be set as int, because global reduction is not supported yet.
  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axis is left in the result as dimension with size one.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.argmax_channel(data=None, out=None, name=None, **kwargs)

Returns argmax indices of each channel from the input array.

The result will be an NDArray of shape (num_channel,).

In case of multiple occurrences of the maximum values, the indices corresponding to the first occurrence are returned.

Examples:

x = [[ 0.,  1.,  2.],
     [ 3.,  4.,  5.]]

argmax_channel(x) = [ 2.,  2.]

Defined in src/operator/tensor/broadcast_reduce_op_index.cc:L97

Parameters:
  • data (NDArray) – The input array
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.argmin(data=None, axis=_Null, keepdims=_Null, out=None, name=None, **kwargs)

Returns indices of the minimum values along an axis.

In the case of multiple occurrences of minimum values, the indices corresponding to the first occurrence are returned.

Examples:

x = [[ 0.,  1.,  2.],
     [ 3.,  4.,  5.]]

// argmin along axis 0
argmin(x, axis=0) = [ 0.,  0.,  0.]

// argmin along axis 1
argmin(x, axis=1) = [ 0.,  0.]

// argmin along axis 1 keeping same dims as an input array
argmin(x, axis=1, keepdims=True) = [[ 0.],
                                    [ 0.]]

Defined in src/operator/tensor/broadcast_reduce_op_index.cc:L77

Parameters:
  • data (NDArray) – The input
  • axis (int or None, optional, default='None') – The axis along which to perform the reduction. Negative values means indexing from right to left. Requires axis to be set as int, because global reduction is not supported yet.
  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axis is left in the result as dimension with size one.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.argsort(data=None, axis=_Null, is_ascend=_Null, dtype=_Null, out=None, name=None, **kwargs)

Returns the indices that would sort an input array along the given axis.

This function performs sorting along the given axis and returns an array of indices having same shape as an input array that index data in sorted order.

Examples:

x = [[ 0.3,  0.2,  0.4],
     [ 0.1,  0.3,  0.2]]

// sort along axis -1
argsort(x) = [[ 1.,  0.,  2.],
              [ 0.,  2.,  1.]]

// sort along axis 0
argsort(x, axis=0) = [[ 1.,  0.,  1.]
                      [ 0.,  1.,  0.]]

// flatten and then sort
argsort(x) = [ 3.,  1.,  5.,  0.,  4.,  2.]

Defined in src/operator/tensor/ordering_op.cc:L177

Parameters:
  • data (NDArray) – The input array
  • axis (int or None, optional, default='-1') – Axis along which to sort the input tensor. If not given, the flattened array is used. Default is -1.
  • is_ascend (boolean, optional, default=1) – Whether to sort in ascending or descending order.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'uint8'},optional, default='float32') – DType of the output indices. It is only valid when ret_typ is “indices” or “both”. An error will be raised if the selected data type cannot precisely represent the indices.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.batch_dot(lhs=None, rhs=None, transpose_a=_Null, transpose_b=_Null, forward_stype=_Null, out=None, name=None, **kwargs)

Batchwise dot product.

batch_dot is used to compute dot product of x and y when x and y are data in batch, namely 3D arrays in shape of (batch_size, :, :).

For example, given x with shape (batch_size, n, m) and y with shape (batch_size, m, k), the result array will have shape (batch_size, n, k), which is computed by:

batch_dot(x,y)[i,:,:] = dot(x[i,:,:], y[i,:,:])

Defined in src/operator/tensor/dot.cc:L125

Parameters:
  • lhs (NDArray) – The first input
  • rhs (NDArray) – The second input
  • transpose_a (boolean, optional, default=0) – If true then transpose the first input before dot.
  • transpose_b (boolean, optional, default=0) – If true then transpose the second input before dot.
  • forward_stype ({None, 'csr', 'default', 'row_sparse'},optional, default='None') – The desired storage type of the forward output given by user, if thecombination of input storage types and this hint does not matchany implemented ones, the dot operator will perform fallback operationand still produce an output of the desired storage type.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.batch_take(a=None, indices=None, out=None, name=None, **kwargs)

Takes elements from a data batch.

Note

batch_take is deprecated. Use pick instead.

Given an input array of shape (d0, d1) and indices of shape (i0,), the result will be an output array of shape (i0,) with:

output[i] = input[i, indices[i]]

Examples:

x = [[ 1.,  2.],
     [ 3.,  4.],
     [ 5.,  6.]]

// takes elements with specified indices
batch_take(x, [0,1,0]) = [ 1.  4.  5.]

Defined in src/operator/tensor/indexing_op.cc:L753

Parameters:
  • a (NDArray) – The input array
  • indices (NDArray) – The index array
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_add(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise sum of the input arrays with broadcasting.

broadcast_plus is an alias to the function broadcast_add.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_add(x, y) = [[ 1.,  1.,  1.],
                       [ 2.,  2.,  2.]]

broadcast_plus(x, y) = [[ 1.,  1.,  1.],
                        [ 2.,  2.,  2.]]

Supported sparse operations:

broadcast_add(csr, dense(1D)) = dense broadcast_add(dense(1D), csr) = dense

Defined in src/operator/tensor/elemwise_binary_broadcast_op_basic.cc:L58

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_axes(data=None, axis=_Null, size=_Null, out=None, name=None, **kwargs)

Broadcasts the input array over particular axes.

Broadcasting is allowed on axes with size 1, such as from (2,1,3,1) to (2,8,3,9). Elements will be duplicated on the broadcasted axes.

Example:

// given x of shape (1,2,1)
x = [[[ 1.],
      [ 2.]]]

// broadcast x on on axis 2
broadcast_axis(x, axis=2, size=3) = [[[ 1.,  1.,  1.],
                                      [ 2.,  2.,  2.]]]
// broadcast x on on axes 0 and 2
broadcast_axis(x, axis=(0,2), size=(2,3)) = [[[ 1.,  1.,  1.],
                                              [ 2.,  2.,  2.]],
                                             [[ 1.,  1.,  1.],
                                              [ 2.,  2.,  2.]]]

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L238

Parameters:
  • data (NDArray) – The input
  • axis (Shape(tuple), optional, default=[]) – The axes to perform the broadcasting.
  • size (Shape(tuple), optional, default=[]) – Target sizes of the broadcasting axes.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_axis(data=None, axis=_Null, size=_Null, out=None, name=None, **kwargs)

Broadcasts the input array over particular axes.

Broadcasting is allowed on axes with size 1, such as from (2,1,3,1) to (2,8,3,9). Elements will be duplicated on the broadcasted axes.

Example:

// given x of shape (1,2,1)
x = [[[ 1.],
      [ 2.]]]

// broadcast x on on axis 2
broadcast_axis(x, axis=2, size=3) = [[[ 1.,  1.,  1.],
                                      [ 2.,  2.,  2.]]]
// broadcast x on on axes 0 and 2
broadcast_axis(x, axis=(0,2), size=(2,3)) = [[[ 1.,  1.,  1.],
                                              [ 2.,  2.,  2.]],
                                             [[ 1.,  1.,  1.],
                                              [ 2.,  2.,  2.]]]

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L238

Parameters:
  • data (NDArray) – The input
  • axis (Shape(tuple), optional, default=[]) – The axes to perform the broadcasting.
  • size (Shape(tuple), optional, default=[]) – Target sizes of the broadcasting axes.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_div(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise division of the input arrays with broadcasting.

Example:

x = [[ 6.,  6.,  6.],
     [ 6.,  6.,  6.]]

y = [[ 2.],
     [ 3.]]

broadcast_div(x, y) = [[ 3.,  3.,  3.],
                       [ 2.,  2.,  2.]]

Supported sparse operations:

broadcast_div(csr, dense(1D)) = csr

Defined in src/operator/tensor/elemwise_binary_broadcast_op_basic.cc:L187

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_equal(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise equal to (==) comparison operation with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_equal(x, y) = [[ 0.,  0.,  0.],
                         [ 1.,  1.,  1.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L46

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_greater(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise greater than (>) comparison operation with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_greater(x, y) = [[ 1.,  1.,  1.],
                           [ 0.,  0.,  0.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L82

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_greater_equal(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise greater than or equal to (>=) comparison operation with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_greater_equal(x, y) = [[ 1.,  1.,  1.],
                                 [ 1.,  1.,  1.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L100

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_hypot(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the hypotenuse of a right angled triangle, given its “legs” with broadcasting.

It is equivalent to doing \(sqrt(x_1^2 + x_2^2)\).

Example:

x = [[ 3.,  3.,  3.]]

y = [[ 4.],
     [ 4.]]

broadcast_hypot(x, y) = [[ 5.,  5.,  5.],
                         [ 5.,  5.,  5.]]

z = [[ 0.],
     [ 4.]]

broadcast_hypot(x, z) = [[ 3.,  3.,  3.],
                         [ 5.,  5.,  5.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_extended.cc:L156

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_lesser(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise lesser than (<) comparison operation with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_lesser(x, y) = [[ 0.,  0.,  0.],
                          [ 0.,  0.,  0.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L118

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_lesser_equal(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise lesser than or equal to (<=) comparison operation with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_lesser_equal(x, y) = [[ 0.,  0.,  0.],
                                [ 1.,  1.,  1.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L136

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_like(lhs=None, rhs=None, lhs_axes=_Null, rhs_axes=_Null, out=None, name=None, **kwargs)

Broadcasts lhs to have the same shape as rhs.

Broadcasting is a mechanism that allows NDArrays to perform arithmetic operations with arrays of different shapes efficiently without creating multiple copies of arrays. Also see, Broadcasting for more explanation.

Broadcasting is allowed on axes with size 1, such as from (2,1,3,1) to (2,8,3,9). Elements will be duplicated on the broadcasted axes.

For example:

broadcast_like([[1,2,3]], [[5,6,7],[7,8,9]]) = [[ 1.,  2.,  3.],
                                                [ 1.,  2.,  3.]])

broadcast_like([9], [1,2,3,4,5], lhs_axes=(0,), rhs_axes=(-1,)) = [9,9,9,9,9]

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L315

Parameters:
  • lhs (NDArray) – First input.
  • rhs (NDArray) – Second input.
  • lhs_axes (Shape or None, optional, default=None) – Axes to perform broadcast on in the first input array
  • rhs_axes (Shape or None, optional, default=None) – Axes to copy from the second input array
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_logical_and(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise logical and with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_logical_and(x, y) = [[ 0.,  0.,  0.],
                               [ 1.,  1.,  1.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L154

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_logical_or(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise logical or with broadcasting.

Example:

x = [[ 1.,  1.,  0.],
     [ 1.,  1.,  0.]]

y = [[ 1.],
     [ 0.]]

broadcast_logical_or(x, y) = [[ 1.,  1.,  1.],
                              [ 1.,  1.,  0.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L172

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_logical_xor(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise logical xor with broadcasting.

Example:

x = [[ 1.,  1.,  0.],
     [ 1.,  1.,  0.]]

y = [[ 1.],
     [ 0.]]

broadcast_logical_xor(x, y) = [[ 0.,  0.,  1.],
                               [ 1.,  1.,  0.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L190

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_maximum(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise maximum of the input arrays with broadcasting.

This function compares two input arrays and returns a new array having the element-wise maxima.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_maximum(x, y) = [[ 1.,  1.,  1.],
                           [ 1.,  1.,  1.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_extended.cc:L80

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_minimum(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise minimum of the input arrays with broadcasting.

This function compares two input arrays and returns a new array having the element-wise minima.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_maximum(x, y) = [[ 0.,  0.,  0.],
                           [ 1.,  1.,  1.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_extended.cc:L115

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_minus(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise difference of the input arrays with broadcasting.

broadcast_minus is an alias to the function broadcast_sub.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_sub(x, y) = [[ 1.,  1.,  1.],
                       [ 0.,  0.,  0.]]

broadcast_minus(x, y) = [[ 1.,  1.,  1.],
                         [ 0.,  0.,  0.]]

Supported sparse operations:

broadcast_sub/minus(csr, dense(1D)) = dense broadcast_sub/minus(dense(1D), csr) = dense

Defined in src/operator/tensor/elemwise_binary_broadcast_op_basic.cc:L106

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_mod(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise modulo of the input arrays with broadcasting.

Example:

x = [[ 8.,  8.,  8.],
     [ 8.,  8.,  8.]]

y = [[ 2.],
     [ 3.]]

broadcast_mod(x, y) = [[ 0.,  0.,  0.],
                       [ 2.,  2.,  2.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_basic.cc:L222

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_mul(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise product of the input arrays with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_mul(x, y) = [[ 0.,  0.,  0.],
                       [ 1.,  1.,  1.]]

Supported sparse operations:

broadcast_mul(csr, dense(1D)) = csr

Defined in src/operator/tensor/elemwise_binary_broadcast_op_basic.cc:L146

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_not_equal(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns the result of element-wise not equal to (!=) comparison operation with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_not_equal(x, y) = [[ 1.,  1.,  1.],
                             [ 0.,  0.,  0.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_logic.cc:L64

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_plus(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise sum of the input arrays with broadcasting.

broadcast_plus is an alias to the function broadcast_add.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_add(x, y) = [[ 1.,  1.,  1.],
                       [ 2.,  2.,  2.]]

broadcast_plus(x, y) = [[ 1.,  1.,  1.],
                        [ 2.,  2.,  2.]]

Supported sparse operations:

broadcast_add(csr, dense(1D)) = dense broadcast_add(dense(1D), csr) = dense

Defined in src/operator/tensor/elemwise_binary_broadcast_op_basic.cc:L58

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_power(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns result of first array elements raised to powers from second array, element-wise with broadcasting.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_power(x, y) = [[ 2.,  2.,  2.],
                         [ 4.,  4.,  4.]]

Defined in src/operator/tensor/elemwise_binary_broadcast_op_extended.cc:L45

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_sub(lhs=None, rhs=None, out=None, name=None, **kwargs)

Returns element-wise difference of the input arrays with broadcasting.

broadcast_minus is an alias to the function broadcast_sub.

Example:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

y = [[ 0.],
     [ 1.]]

broadcast_sub(x, y) = [[ 1.,  1.,  1.],
                       [ 0.,  0.,  0.]]

broadcast_minus(x, y) = [[ 1.,  1.,  1.],
                         [ 0.,  0.,  0.]]

Supported sparse operations:

broadcast_sub/minus(csr, dense(1D)) = dense broadcast_sub/minus(dense(1D), csr) = dense

Defined in src/operator/tensor/elemwise_binary_broadcast_op_basic.cc:L106

Parameters:
  • lhs (NDArray) – First input to the function
  • rhs (NDArray) – Second input to the function
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.broadcast_to(data=None, shape=_Null, out=None, name=None, **kwargs)

Broadcasts the input array to a new shape.

Broadcasting is a mechanism that allows NDArrays to perform arithmetic operations with arrays of different shapes efficiently without creating multiple copies of arrays. Also see, Broadcasting for more explanation.

Broadcasting is allowed on axes with size 1, such as from (2,1,3,1) to (2,8,3,9). Elements will be duplicated on the broadcasted axes.

For example:

broadcast_to([[1,2,3]], shape=(2,3)) = [[ 1.,  2.,  3.],
                                        [ 1.,  2.,  3.]])

The dimension which you do not want to change can also be kept as 0 which means copy the original value. So with shape=(2,0), we will obtain the same result as in the above example.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L262

Parameters:
  • data (NDArray) – The input
  • shape (Shape(tuple), optional, default=[]) – The shape of the desired array. We can set the dim to zero if it’s same as the original. E.g A = broadcast_to(B, shape=(10, 0, 0)) has the same meaning as A = broadcast_axis(B, axis=0, size=10).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.cast(data=None, dtype=_Null, out=None, name=None, **kwargs)

Casts all elements of the input to a new type.

Note

Cast is deprecated. Use cast instead.

Example:

cast([0.9, 1.3], dtype='int32') = [0, 1]
cast([1e20, 11.1], dtype='float16') = [inf, 11.09375]
cast([300, 11.1, 10.9, -1, -3], dtype='uint8') = [44, 11, 10, 255, 253]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L634

Parameters:
  • data (NDArray) – The input.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8'}, required) – Output data type.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.cast_storage(data=None, stype=_Null, out=None, name=None, **kwargs)

Casts tensor storage type to the new type.

When an NDArray with default storage type is cast to csr or row_sparse storage, the result is compact, which means:

  • for csr, zero values will not be retained
  • for row_sparse, row slices of all zeros will not be retained

The storage type of cast_storage output depends on stype parameter:

  • cast_storage(csr, ‘default’) = default
  • cast_storage(row_sparse, ‘default’) = default
  • cast_storage(default, ‘csr’) = csr
  • cast_storage(default, ‘row_sparse’) = row_sparse
  • cast_storage(csr, ‘csr’) = csr
  • cast_storage(row_sparse, ‘row_sparse’) = row_sparse

Example:

dense = [[ 0.,  1.,  0.],
         [ 2.,  0.,  3.],
         [ 0.,  0.,  0.],
         [ 0.,  0.,  0.]]

# cast to row_sparse storage type
rsp = cast_storage(dense, 'row_sparse')
rsp.indices = [0, 1]
rsp.values = [[ 0.,  1.,  0.],
              [ 2.,  0.,  3.]]

# cast to csr storage type
csr = cast_storage(dense, 'csr')
csr.indices = [1, 0, 2]
csr.values = [ 1.,  2.,  3.]
csr.indptr = [0, 1, 3, 3, 3]

Defined in src/operator/tensor/cast_storage.cc:L71

Parameters:
  • data (NDArray) – The input.
  • stype ({'csr', 'default', 'row_sparse'}, required) – Output storage type.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.cbrt(data=None, out=None, name=None, **kwargs)

Returns element-wise cube-root value of the input.

\[cbrt(x) = \sqrt[3]{x}\]

Example:

cbrt([1, 8, -125]) = [1, 2, -5]

The storage type of cbrt output depends upon the input storage type:

  • cbrt(default) = default
  • cbrt(row_sparse) = row_sparse
  • cbrt(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L950

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ceil(data=None, out=None, name=None, **kwargs)

Returns element-wise ceiling of the input.

The ceil of the scalar x is the smallest integer i, such that i >= x.

Example:

ceil([-2.1, -1.9, 1.5, 1.9, 2.1]) = [-2., -1.,  2.,  2.,  3.]

The storage type of ceil output depends upon the input storage type:

  • ceil(default) = default
  • ceil(row_sparse) = row_sparse
  • ceil(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L786

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.choose_element_0index(data=None, index=None, axis=_Null, keepdims=_Null, mode=_Null, out=None, name=None, **kwargs)

Picks elements from an input array according to the input indices along the given axis.

Given an input array of shape (d0, d1) and indices of shape (i0,), the result will be an output array of shape (i0,) with:

output[i] = input[i, indices[i]]

By default, if any index mentioned is too large, it is replaced by the index that addresses the last element along an axis (the clip mode).

This function supports n-dimensional input and (n-1)-dimensional indices arrays.

Examples:

x = [[ 1.,  2.],
     [ 3.,  4.],
     [ 5.,  6.]]

// picks elements with specified indices along axis 0
pick(x, y=[0,1], 0) = [ 1.,  4.]

// picks elements with specified indices along axis 1
pick(x, y=[0,1,0], 1) = [ 1.,  4.,  5.]

y = [[ 1.],
     [ 0.],
     [ 2.]]

// picks elements with specified indices along axis 1 using 'wrap' mode
// to place indicies that would normally be out of bounds
pick(x, y=[2,-1,-2], 1, mode='wrap') = [ 1.,  4.,  5.]

y = [[ 1.],
     [ 0.],
     [ 2.]]

// picks elements with specified indices along axis 1 and dims are maintained
pick(x,y, 1, keepdims=True) = [[ 2.],
                               [ 3.],
                               [ 6.]]

Defined in src/operator/tensor/broadcast_reduce_op_index.cc:L154

Parameters:
  • data (NDArray) – The input array
  • index (NDArray) – The index array
  • axis (int or None, optional, default='-1') – int or None. The axis to picking the elements. Negative values means indexing from right to left. If is None, the elements in the index w.r.t the flattened input will be picked.
  • keepdims (boolean, optional, default=0) – If true, the axis where we pick the elements is left in the result as dimension with size one.
  • mode ({'clip', 'wrap'},optional, default='clip') – Specify how out-of-bound indices behave. Default is “clip”. “clip” means clip to the range. So, if all indices mentioned are too large, they are replaced by the index that addresses the last element along an axis. “wrap” means to wrap around.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.clip(data=None, a_min=_Null, a_max=_Null, out=None, name=None, **kwargs)

Clips (limits) the values in an array.

Given an interval, values outside the interval are clipped to the interval edges. Clipping x between a_min and a_x would be:

clip(x, a_min, a_max) = max(min(x, a_max), a_min))

Example:

x = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

clip(x,1,8) = [ 1.,  1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  8.]

The storage type of clip output depends on storage types of inputs and the a_min, a_max parameter values:

  • clip(default) = default
  • clip(row_sparse, a_min <= 0, a_max >= 0) = row_sparse
  • clip(csr, a_min <= 0, a_max >= 0) = csr
  • clip(row_sparse, a_min < 0, a_max < 0) = default
  • clip(row_sparse, a_min > 0, a_max > 0) = default
  • clip(csr, a_min < 0, a_max < 0) = csr
  • clip(csr, a_min > 0, a_max > 0) = csr

Defined in src/operator/tensor/matrix_op.cc:L723

Parameters:
  • data (NDArray) – Input array.
  • a_min (float, required) – Minimum value
  • a_max (float, required) – Maximum value
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.concat(*data, **kwargs)

Joins input arrays along a given axis.

Note

Concat is deprecated. Use concat instead.

The dimensions of the input arrays should be the same except the axis along which they will be concatenated. The dimension of the output array along the concatenated axis will be equal to the sum of the corresponding dimensions of the input arrays.

The storage type of concat output depends on storage types of inputs

  • concat(csr, csr, ..., csr, dim=0) = csr
  • otherwise, concat generates output with default storage

Example:

x = [[1,1],[2,2]]
y = [[3,3],[4,4],[5,5]]
z = [[6,6], [7,7],[8,8]]

concat(x,y,z,dim=0) = [[ 1.,  1.],
                       [ 2.,  2.],
                       [ 3.,  3.],
                       [ 4.,  4.],
                       [ 5.,  5.],
                       [ 6.,  6.],
                       [ 7.,  7.],
                       [ 8.,  8.]]

Note that you cannot concat x,y,z along dimension 1 since dimension
0 is not the same for all the input arrays.

concat(y,z,dim=1) = [[ 3.,  3.,  6.,  6.],
                      [ 4.,  4.,  7.,  7.],
                      [ 5.,  5.,  8.,  8.]]

Defined in src/operator/nn/concat.cc:L371

Parameters:
  • data (NDArray[]) – List of arrays to concatenate
  • dim (int, optional, default='1') – the dimension to be concated.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.cos(data=None, out=None, name=None, **kwargs)

Computes the element-wise cosine of the input array.

The input should be in radians (\(2\pi\) rad equals 360 degrees).

\[cos([0, \pi/4, \pi/2]) = [1, 0.707, 0]\]

The storage type of cos output is always dense

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L89

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.cosh(data=None, out=None, name=None, **kwargs)

Returns the hyperbolic cosine of the input array, computed element-wise.

\[cosh(x) = 0.5\times(exp(x) + exp(-x))\]

The storage type of cosh output is always dense

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L272

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.crop(data=None, begin=_Null, end=_Null, step=_Null, out=None, name=None, **kwargs)

Slices a region of the array.

Note

crop is deprecated. Use slice instead.

This function returns a sliced array between the indices given by begin and end with the corresponding step.

For an input array of shape=(d_0, d_1, ..., d_n-1), slice operation with begin=(b_0, b_1...b_m-1), end=(e_0, e_1, ..., e_m-1), and step=(s_0, s_1, ..., s_m-1), where m <= n, results in an array with the shape (|e_0-b_0|/|s_0|, ..., |e_m-1-b_m-1|/|s_m-1|, d_m, ..., d_n-1).

The resulting array’s k-th dimension contains elements from the k-th dimension of the input array starting from index b_k (inclusive) with step s_k until reaching e_k (exclusive).

If the k-th elements are None in the sequence of begin, end, and step, the following rule will be used to set default values. If s_k is None, set s_k=1. If s_k > 0, set b_k=0, e_k=d_k; else, set b_k=d_k-1, e_k=-1.

The storage type of slice output depends on storage types of inputs

  • slice(csr) = csr
  • otherwise, slice generates output with default storage

Note

When input data storage type is csr, it only supports step=(), or step=(None,), or step=(1,) to generate a csr output. For other step parameter values, it falls back to slicing a dense tensor.

Example:

x = [[  1.,   2.,   3.,   4.],
     [  5.,   6.,   7.,   8.],
     [  9.,  10.,  11.,  12.]]

slice(x, begin=(0,1), end=(2,4)) = [[ 2.,  3.,  4.],
                                   [ 6.,  7.,  8.]]
slice(x, begin=(None, 0), end=(None, 3), step=(-1, 2)) = [[9., 11.],
                                                          [5.,  7.],
                                                          [1.,  3.]]

Defined in src/operator/tensor/matrix_op.cc:L506

Parameters:
  • data (NDArray) – Source input
  • begin (Shape(tuple), required) – starting indices for the slice operation, supports negative indices.
  • end (Shape(tuple), required) – ending indices for the slice operation, supports negative indices.
  • step (Shape(tuple), optional, default=[]) – step for the slice operation, supports negative values.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ctc_loss(data=None, label=None, data_lengths=None, label_lengths=None, use_data_lengths=_Null, use_label_lengths=_Null, blank_label=_Null, out=None, name=None, **kwargs)

Connectionist Temporal Classification Loss.

Note

The existing alias contrib_CTCLoss is deprecated.

The shapes of the inputs and outputs:

  • data: (sequence_length, batch_size, alphabet_size)
  • label: (batch_size, label_sequence_length)
  • out: (batch_size)

The data tensor consists of sequences of activation vectors (without applying softmax), with i-th channel in the last dimension corresponding to i-th label for i between 0 and alphabet_size-1 (i.e always 0-indexed). Alphabet size should include one additional value reserved for blank label. When blank_label is "first", the 0-th channel is be reserved for activation of blank label, or otherwise if it is “last”, (alphabet_size-1)-th channel should be reserved for blank label.

label is an index matrix of integers. When blank_label is "first", the value 0 is then reserved for blank label, and should not be passed in this matrix. Otherwise, when blank_label is "last", the value (alphabet_size-1) is reserved for blank label.

If a sequence of labels is shorter than label_sequence_length, use the special padding value at the end of the sequence to conform it to the correct length. The padding value is 0 when blank_label is "first", and -1 otherwise.

For example, suppose the vocabulary is [a, b, c], and in one batch we have three sequences ‘ba’, ‘cbb’, and ‘abac’. When blank_label is "first", we can index the labels as {‘a’: 1, ‘b’: 2, ‘c’: 3}, and we reserve the 0-th channel for blank label in data tensor. The resulting label tensor should be padded to be:

[[2, 1, 0, 0], [3, 2, 2, 0], [1, 2, 1, 3]]

When blank_label is "last", we can index the labels as {‘a’: 0, ‘b’: 1, ‘c’: 2}, and we reserve the channel index 3 for blank label in data tensor. The resulting label tensor should be padded to be:

[[1, 0, -1, -1], [2, 1, 1, -1], [0, 1, 0, 2]]

out is a list of CTC loss values, one per example in the batch.

See Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks, A. Graves et al. for more information on the definition and the algorithm.

Defined in src/operator/nn/ctc_loss.cc:L100

Parameters:
  • data (NDArray) – Input ndarray
  • label (NDArray) – Ground-truth labels for the loss.
  • data_lengths (NDArray) – Lengths of data for each of the samples. Only required when use_data_lengths is true.
  • label_lengths (NDArray) – Lengths of labels for each of the samples. Only required when use_label_lengths is true.
  • use_data_lengths (boolean, optional, default=0) – Whether the data lenghts are decided by data_lengths. If false, the lengths are equal to the max sequence length.
  • use_label_lengths (boolean, optional, default=0) – Whether the label lenghts are decided by label_lengths, or derived from padding_mask. If false, the lengths are derived from the first occurrence of the value of padding_mask. The value of padding_mask is 0 when first CTC label is reserved for blank, and -1 when last label is reserved for blank. See blank_label.
  • blank_label ({'first', 'last'},optional, default='first') – Set the label that is reserved for blank label.If “first”, 0-th label is reserved, and label values for tokens in the vocabulary are between 1 and alphabet_size-1, and the padding mask is -1. If “last”, last label value alphabet_size-1 is reserved for blank label instead, and label values for tokens in the vocabulary are between 0 and alphabet_size-2, and the padding mask is 0.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.degrees(data=None, out=None, name=None, **kwargs)

Converts each element of the input array from radians to degrees.

\[degrees([0, \pi/2, \pi, 3\pi/2, 2\pi]) = [0, 90, 180, 270, 360]\]

The storage type of degrees output depends upon the input storage type:

  • degrees(default) = default
  • degrees(row_sparse) = row_sparse
  • degrees(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L219

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.depth_to_space(data=None, block_size=_Null, out=None, name=None, **kwargs)

Rearranges(permutes) data from depth into blocks of spatial data. Similar to ONNX DepthToSpace operator: https://github.com/onnx/onnx/blob/master/docs/Operators.md#DepthToSpace. The output is a new tensor where the values from depth dimension are moved in spatial blocks to height and width dimension. The reverse of this operation is space_to_depth.

\[\begin{split}\begin{gather*} x \prime = reshape(x, [N, block\_size, block\_size, C / (block\_size ^ 2), H * block\_size, W * block\_size]) \\ x \prime \prime = transpose(x \prime, [0, 3, 4, 1, 5, 2]) \\ y = reshape(x \prime \prime, [N, C / (block\_size ^ 2), H * block\_size, W * block\_size]) \end{gather*}\end{split}\]

where \(x\) is an input tensor with default layout as \([N, C, H, W]\): [batch, channels, height, width] and \(y\) is the output tensor of layout \([N, C / (block\_size ^ 2), H * block\_size, W * block\_size]\)

Example:

x = [[[[0, 1, 2],
       [3, 4, 5]],
      [[6, 7, 8],
       [9, 10, 11]],
      [[12, 13, 14],
       [15, 16, 17]],
      [[18, 19, 20],
       [21, 22, 23]]]]

depth_to_space(x, 2) = [[[[0, 6, 1, 7, 2, 8],
                          [12, 18, 13, 19, 14, 20],
                          [3, 9, 4, 10, 5, 11],
                          [15, 21, 16, 22, 17, 23]]]]

Defined in src/operator/tensor/matrix_op.cc:L1050

Parameters:
  • data (NDArray) – Input ndarray
  • block_size (int, required) – Blocks of [block_size. block_size] are moved
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.diag(data=None, k=_Null, axis1=_Null, axis2=_Null, out=None, name=None, **kwargs)

Extracts a diagonal or constructs a diagonal array.

diag‘s behavior depends on the input array dimensions:

  • 1-D arrays: constructs a 2-D array with the input as its diagonal, all other elements are zero.

  • N-D arrays: extracts the diagonals of the sub-arrays with axes specified by axis1 and axis2. The output shape would be decided by removing the axes numbered axis1 and axis2 from the input shape and appending to the result a new axis with the size of the diagonals in question.

    For example, when the input shape is (2, 3, 4, 5), axis1 and axis2 are 0 and 2 respectively and k is 0, the resulting shape would be (3, 5, 2).

Examples:

x = [[1, 2, 3],
     [4, 5, 6]]

diag(x) = [1, 5]

diag(x, k=1) = [2, 6]

diag(x, k=-1) = [4]

x = [1, 2, 3]

diag(x) = [[1, 0, 0],
           [0, 2, 0],
           [0, 0, 3]]

diag(x, k=1) = [[0, 1, 0],
                [0, 0, 2],
                [0, 0, 0]]

diag(x, k=-1) = [[0, 0, 0],
                 [1, 0, 0],
                 [0, 2, 0]]

x = [[[1, 2],
      [3, 4]],

     [[5, 6],
      [7, 8]]]

diag(x) = [[1, 7],
           [2, 8]]

diag(x, k=1) = [[3],
                [4]]

diag(x, axis1=-2, axis2=-1) = [[1, 4],
                               [5, 8]]

Defined in src/operator/tensor/diag_op.cc:L87

Parameters:
  • data (NDArray) – Input ndarray
  • k (int, optional, default='0') – Diagonal in question. The default is 0. Use k>0 for diagonals above the main diagonal, and k<0 for diagonals below the main diagonal. If input has shape (S0 S1) k must be between -S0 and S1
  • axis1 (int, optional, default='0') – The first axis of the sub-arrays of interest. Ignored when the input is a 1-D array.
  • axis2 (int, optional, default='1') – The second axis of the sub-arrays of interest. Ignored when the input is a 1-D array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.dot(lhs=None, rhs=None, transpose_a=_Null, transpose_b=_Null, forward_stype=_Null, out=None, name=None, **kwargs)

Dot product of two arrays.

dot‘s behavior depends on the input array dimensions:

  • 1-D arrays: inner product of vectors

  • 2-D arrays: matrix multiplication

  • N-D arrays: a sum product over the last axis of the first input and the first axis of the second input

    For example, given 3-D x with shape (n,m,k) and y with shape (k,r,s), the result array will have shape (n,m,r,s). It is computed by:

    dot(x,y)[i,j,a,b] = sum(x[i,j,:]*y[:,a,b])
    

    Example:

    x = reshape([0,1,2,3,4,5,6,7], shape=(2,2,2))
    y = reshape([7,6,5,4,3,2,1,0], shape=(2,2,2))
    dot(x,y)[0,0,1,1] = 0
    sum(x[0,0,:]*y[:,1,1]) = 0
    

The storage type of dot output depends on storage types of inputs, transpose option and forward_stype option for output storage type. Implemented sparse operations include:

  • dot(default, default, transpose_a=True/False, transpose_b=True/False) = default
  • dot(csr, default, transpose_a=True) = default
  • dot(csr, default, transpose_a=True) = row_sparse
  • dot(csr, default) = default
  • dot(csr, row_sparse) = default
  • dot(default, csr) = csr (CPU only)
  • dot(default, csr, forward_stype=’default’) = default
  • dot(default, csr, transpose_b=True, forward_stype=’default’) = default

If the combination of input storage types and forward_stype does not match any of the above patterns, dot will fallback and generate output with default storage.

Note

If the storage type of the lhs is “csr”, the storage type of gradient w.r.t rhs will be “row_sparse”. Only a subset of optimizers support sparse gradients, including SGD, AdaGrad and Adam. Note that by default lazy updates is turned on, which may perform differently from standard updates. For more details, please check the Optimization API at: /api/python/optimization/optimization.html

Defined in src/operator/tensor/dot.cc:L77

Parameters:
  • lhs (NDArray) – The first input
  • rhs (NDArray) – The second input
  • transpose_a (boolean, optional, default=0) – If true then transpose the first input before dot.
  • transpose_b (boolean, optional, default=0) – If true then transpose the second input before dot.
  • forward_stype ({None, 'csr', 'default', 'row_sparse'},optional, default='None') – The desired storage type of the forward output given by user, if thecombination of input storage types and this hint does not matchany implemented ones, the dot operator will perform fallback operationand still produce an output of the desired storage type.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.elemwise_add(lhs=None, rhs=None, out=None, name=None, **kwargs)

Adds arguments element-wise.

The storage type of elemwise_add output depends on storage types of inputs

  • elemwise_add(row_sparse, row_sparse) = row_sparse
  • elemwise_add(csr, csr) = csr
  • elemwise_add(default, csr) = default
  • elemwise_add(csr, default) = default
  • elemwise_add(default, rsp) = default
  • elemwise_add(rsp, default) = default
  • otherwise, elemwise_add generates output with default storage
Parameters:
  • lhs (NDArray) – first input
  • rhs (NDArray) – second input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

Example

>>> x = mx.nd.array([1, 2, 3, 4])
>>> y = mx.nd.array([1.1, 2.1, 3.1, 4.1])
>>> mx.nd.elemwise_add(x, y).asnumpy()
array([ 2.0999999 ,  4.0999999 ,  6.0999999 ,  8.10000038], dtype=float32)
mxnet.ndarray.elemwise_div(lhs=None, rhs=None, out=None, name=None, **kwargs)

Divides arguments element-wise.

The storage type of elemwise_div output is always dense

Parameters:
  • lhs (NDArray) – first input
  • rhs (NDArray) – second input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.elemwise_mul(lhs=None, rhs=None, out=None, name=None, **kwargs)

Multiplies arguments element-wise.

The storage type of elemwise_mul output depends on storage types of inputs

  • elemwise_mul(default, default) = default
  • elemwise_mul(row_sparse, row_sparse) = row_sparse
  • elemwise_mul(default, row_sparse) = row_sparse
  • elemwise_mul(row_sparse, default) = row_sparse
  • elemwise_mul(csr, csr) = csr
  • otherwise, elemwise_mul generates output with default storage
Parameters:
  • lhs (NDArray) – first input
  • rhs (NDArray) – second input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.elemwise_sub(lhs=None, rhs=None, out=None, name=None, **kwargs)

Subtracts arguments element-wise.

The storage type of elemwise_sub output depends on storage types of inputs

  • elemwise_sub(row_sparse, row_sparse) = row_sparse
  • elemwise_sub(csr, csr) = csr
  • elemwise_sub(default, csr) = default
  • elemwise_sub(csr, default) = default
  • elemwise_sub(default, rsp) = default
  • elemwise_sub(rsp, default) = default
  • otherwise, elemwise_sub generates output with default storage
Parameters:
  • lhs (NDArray) – first input
  • rhs (NDArray) – second input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.erf(data=None, out=None, name=None, **kwargs)

Returns element-wise gauss error function of the input.

Example:

erf([0, -1., 10.]) = [0., -0.8427, 1.]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L964

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.erfinv(data=None, out=None, name=None, **kwargs)

Returns element-wise inverse gauss error function of the input.

Example:

erfinv([0, 0.5., -1.]) = [0., 0.4769, -inf]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L985

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.exp(data=None, out=None, name=None, **kwargs)

Returns element-wise exponential value of the input.

\[exp(x) = e^x \approx 2.718^x\]

Example:

exp([0, 1, 2]) = [1., 2.71828175, 7.38905621]

The storage type of exp output is always dense

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L1044

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.expand_dims(data=None, axis=_Null, out=None, name=None, **kwargs)

Inserts a new axis of size 1 into the array shape

For example, given x with shape (2,3,4), then expand_dims(x, axis=1) will return a new array with shape (2,1,3,4).

Defined in src/operator/tensor/matrix_op.cc:L416

Parameters:
  • data (NDArray) – Source input
  • axis (int, required) – Position where new axis is to be inserted. Suppose that the input NDArray‘s dimension is ndim, the range of the inserted axis is [-ndim, ndim]
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.expm1(data=None, out=None, name=None, **kwargs)

Returns exp(x) - 1 computed element-wise on the input.

This function provides greater precision than exp(x) - 1 for small values of x.

The storage type of expm1 output depends upon the input storage type:

  • expm1(default) = default
  • expm1(row_sparse) = row_sparse
  • expm1(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L1189

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.fill_element_0index(lhs=None, mhs=None, rhs=None, out=None, name=None, **kwargs)

Fill one element of each line(row for python, column for R/Julia) in lhs according to index indicated by rhs and values indicated by mhs. This function assume rhs uses 0-based index.

Parameters:
  • lhs (NDArray) – Left operand to the function.
  • mhs (NDArray) – Middle operand to the function.
  • rhs (NDArray) – Right operand to the function.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.fix(data=None, out=None, name=None, **kwargs)

Returns element-wise rounded value to the nearest integer towards zero of the input.

Example:

fix([-2.1, -1.9, 1.9, 2.1]) = [-2., -1.,  1., 2.]

The storage type of fix output depends upon the input storage type:

  • fix(default) = default
  • fix(row_sparse) = row_sparse
  • fix(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L843

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.flatten(data=None, out=None, name=None, **kwargs)

Flattens the input array into a 2-D array by collapsing the higher dimensions.

Note

Flatten is deprecated. Use flatten instead.

For an input array with shape (d1, d2, ..., dk), flatten operation reshapes the input array into an output array of shape (d1, d2*...*dk).

Note that the bahavior of this function is different from numpy.ndarray.flatten, which behaves similar to mxnet.ndarray.reshape((-1,)).

Example:

x = [[
    [1,2,3],
    [4,5,6],
    [7,8,9]
],
[    [1,2,3],
    [4,5,6],
    [7,8,9]
]],

flatten(x) = [[ 1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9.],
   [ 1.,  2.,  3.,  4.,  5.,  6.,  7.,  8.,  9.]]

Defined in src/operator/tensor/matrix_op.cc:L291

Parameters:
  • data (NDArray) – Input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.flip(data=None, axis=_Null, out=None, name=None, **kwargs)

Reverses the order of elements along given axis while preserving array shape.

Note: reverse and flip are equivalent. We use reverse in the following examples.

Examples:

x = [[ 0.,  1.,  2.,  3.,  4.],
     [ 5.,  6.,  7.,  8.,  9.]]

reverse(x, axis=0) = [[ 5.,  6.,  7.,  8.,  9.],
                      [ 0.,  1.,  2.,  3.,  4.]]

reverse(x, axis=1) = [[ 4.,  3.,  2.,  1.,  0.],
                      [ 9.,  8.,  7.,  6.,  5.]]

Defined in src/operator/tensor/matrix_op.cc:L898

Parameters:
  • data (NDArray) – Input data array
  • axis (Shape(tuple), required) – The axis which to reverse elements.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.floor(data=None, out=None, name=None, **kwargs)

Returns element-wise floor of the input.

The floor of the scalar x is the largest integer i, such that i <= x.

Example:

floor([-2.1, -1.9, 1.5, 1.9, 2.1]) = [-3., -2.,  1.,  1.,  2.]

The storage type of floor output depends upon the input storage type:

  • floor(default) = default
  • floor(row_sparse) = row_sparse
  • floor(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L805

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ftml_update(weight=None, grad=None, d=None, v=None, z=None, lr=_Null, beta1=_Null, beta2=_Null, epsilon=_Null, t=_Null, wd=_Null, rescale_grad=_Null, clip_grad=_Null, out=None, name=None, **kwargs)

The FTML optimizer described in FTML - Follow the Moving Leader in Deep Learning, available at http://proceedings.mlr.press/v70/zheng17a/zheng17a.pdf.

\[\begin{split}g_t = \nabla J(W_{t-1})\\ v_t = \beta_2 v_{t-1} + (1 - \beta_2) g_t^2\\ d_t = \frac{ 1 - \beta_1^t }{ \eta_t } (\sqrt{ \frac{ v_t }{ 1 - \beta_2^t } } + \epsilon) \sigma_t = d_t - \beta_1 d_{t-1} z_t = \beta_1 z_{ t-1 } + (1 - \beta_1^t) g_t - \sigma_t W_{t-1} W_t = - \frac{ z_t }{ d_t }\end{split}\]

Defined in src/operator/optimizer_op.cc:L638

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • d (NDArray) – Internal state d_t
  • v (NDArray) – Internal state v_t
  • z (NDArray) – Internal state z_t
  • lr (float, required) – Learning rate.
  • beta1 (float, optional, default=0.600000024) – Generally close to 0.5.
  • beta2 (float, optional, default=0.999000013) – Generally close to 1.
  • epsilon (double, optional, default=9.9999999392252903e-09) – Epsilon to prevent div 0.
  • t (int, required) – Number of update.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_grad (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ftrl_update(weight=None, grad=None, z=None, n=None, lr=_Null, lamda1=_Null, beta=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, out=None, name=None, **kwargs)

Update function for Ftrl optimizer. Referenced from Ad Click Prediction: a View from the Trenches, available at http://dl.acm.org/citation.cfm?id=2488200.

It updates the weights using:

rescaled_grad = clip(grad * rescale_grad, clip_gradient)
z += rescaled_grad - (sqrt(n + rescaled_grad**2) - sqrt(n)) * weight / learning_rate
n += rescaled_grad**2
w = (sign(z) * lamda1 - z) / ((beta + sqrt(n)) / learning_rate + wd) * (abs(z) > lamda1)

If w, z and n are all of row_sparse storage type, only the row slices whose indices appear in grad.indices are updated (for w, z and n):

for row in grad.indices:
    rescaled_grad[row] = clip(grad[row] * rescale_grad, clip_gradient)
    z[row] += rescaled_grad[row] - (sqrt(n[row] + rescaled_grad[row]**2) - sqrt(n[row])) * weight[row] / learning_rate
    n[row] += rescaled_grad[row]**2
    w[row] = (sign(z[row]) * lamda1 - z[row]) / ((beta + sqrt(n[row])) / learning_rate + wd) * (abs(z[row]) > lamda1)

Defined in src/operator/optimizer_op.cc:L874

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • z (NDArray) – z
  • n (NDArray) – Square of grad
  • lr (float, required) – Learning rate
  • lamda1 (float, optional, default=0.00999999978) – The L1 regularization coefficient.
  • beta (float, optional, default=1) – Per-Coordinate Learning Rate beta.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.gamma(data=None, out=None, name=None, **kwargs)

Returns the gamma function (extension of the factorial function to the reals), computed element-wise on the input array.

The storage type of gamma output is always dense

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.gammaln(data=None, out=None, name=None, **kwargs)

Returns element-wise log of the absolute value of the gamma function of the input.

The storage type of gammaln output is always dense

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.gather_nd(data=None, indices=None, out=None, name=None, **kwargs)

Gather elements or slices from data and store to a tensor whose shape is defined by indices.

Given data with shape (X_0, X_1, ..., X_{N-1}) and indices with shape (M, Y_0, ..., Y_{K-1}), the output will have shape (Y_0, ..., Y_{K-1}, X_M, ..., X_{N-1}), where M <= N. If M == N, output shape will simply be (Y_0, ..., Y_{K-1}).

The elements in output is defined as follows:

output[y_0, ..., y_{K-1}, x_M, ..., x_{N-1}] = data[indices[0, y_0, ..., y_{K-1}],
                                                    ...,
                                                    indices[M-1, y_0, ..., y_{K-1}],
                                                    x_M, ..., x_{N-1}]

Examples:

data = [[0, 1], [2, 3]]
indices = [[1, 1, 0], [0, 1, 0]]
gather_nd(data, indices) = [2, 3, 0]

data = [[[1, 2], [3, 4]], [[5, 6], [7, 8]]]
indices = [[0, 1], [1, 0]]
gather_nd(data, indices) = [[3, 4], [5, 6]]
Parameters:
  • data (NDArray) – data
  • indices (NDArray) – indices
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.hard_sigmoid(data=None, alpha=_Null, beta=_Null, out=None, name=None, **kwargs)

Computes hard sigmoid of x element-wise.

\[y = max(0, min(1, alpha * x + beta))\]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L133

Parameters:
  • data (NDArray) – The input array.
  • alpha (float, optional, default=0.200000003) – Slope of hard sigmoid
  • beta (float, optional, default=0.5) – Bias of hard sigmoid.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.identity(data=None, out=None, name=None, **kwargs)

Returns a copy of the input.

From:src/operator/tensor/elemwise_unary_op_basic.cc:218

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.khatri_rao(*args, **kwargs)

Computes the Khatri-Rao product of the input matrices.

Given a collection of \(n\) input matrices,

\[A_1 \in \mathbb{R}^{M_1 \times M}, \ldots, A_n \in \mathbb{R}^{M_n \times N},\]

the (column-wise) Khatri-Rao product is defined as the matrix,

\[X = A_1 \otimes \cdots \otimes A_n \in \mathbb{R}^{(M_1 \cdots M_n) \times N},\]

where the \(k\) th column is equal to the column-wise outer product \({A_1}_k \otimes \cdots \otimes {A_n}_k\) where \({A_i}_k\) is the kth column of the ith matrix.

Example:

>>> A = mx.nd.array([[1, -1],
>>>                  [2, -3]])
>>> B = mx.nd.array([[1, 4],
>>>                  [2, 5],
>>>                  [3, 6]])
>>> C = mx.nd.khatri_rao(A, B)
>>> print(C.asnumpy())
[[  1.  -4.]
 [  2.  -5.]
 [  3.  -6.]
 [  2. -12.]
 [  4. -15.]
 [  6. -18.]]

Defined in src/operator/contrib/krprod.cc:L108

Parameters:
  • args (NDArray[]) – Positional input matrices
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_extractdiag(A=None, offset=_Null, out=None, name=None, **kwargs)

Extracts the diagonal entries of a square matrix. Input is a tensor A of dimension n >= 2.

If n=2, then A represents a single square matrix which diagonal elements get extracted as a 1-dimensional tensor.

If n>2, then A represents a batch of square matrices on the trailing two dimensions. The extracted diagonals are returned as an n-1-dimensional tensor.

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix diagonal extraction
A = [[1.0, 2.0],
     [3.0, 4.0]]

extractdiag(A) = [1.0, 4.0]

extractdiag(A, 1) = [2.0]

// Batch matrix diagonal extraction
A = [[[1.0, 2.0],
      [3.0, 4.0]],
     [[5.0, 6.0],
      [7.0, 8.0]]]

extractdiag(A) = [[1.0, 4.0],
                  [5.0, 8.0]]

Defined in src/operator/tensor/la_op.cc:L495

Parameters:
  • A (NDArray) – Tensor of square matrices
  • offset (int, optional, default='0') – Offset of the diagonal versus the main diagonal. 0 corresponds to the main diagonal, a negative/positive value to diagonals below/above the main diagonal.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_extracttrian(A=None, offset=_Null, lower=_Null, out=None, name=None, **kwargs)

Extracts a triangular sub-matrix from a square matrix. Input is a tensor A of dimension n >= 2.

If n=2, then A represents a single square matrix from which a triangular sub-matrix is extracted as a 1-dimensional tensor.

If n>2, then A represents a batch of square matrices on the trailing two dimensions. The extracted triangular sub-matrices are returned as an n-1-dimensional tensor.

The offset and lower parameters determine the triangle to be extracted:

  • When offset = 0 either the lower or upper triangle with respect to the main diagonal is extracted depending on the value of parameter lower.
  • When offset = k > 0 the upper triangle with respect to the k-th diagonal above the main diagonal is extracted.
  • When offset = k < 0 the lower triangle with respect to the k-th diagonal below the main diagonal is extracted.

Note

The operator supports float32 and float64 data types only.

Examples:

// Single triagonal extraction
A = [[1.0, 2.0],
     [3.0, 4.0]]

extracttrian(A) = [1.0, 3.0, 4.0]
extracttrian(A, lower=False) = [1.0, 2.0, 4.0]
extracttrian(A, 1) = [2.0]
extracttrian(A, -1) = [3.0]

// Batch triagonal extraction
A = [[[1.0, 2.0],
      [3.0, 4.0]],
     [[5.0, 6.0],
      [7.0, 8.0]]]

extracttrian(A) = [[1.0, 3.0, 4.0],
                   [5.0, 7.0, 8.0]]

Defined in src/operator/tensor/la_op.cc:L605

Parameters:
  • A (NDArray) – Tensor of square matrices
  • offset (int, optional, default='0') – Offset of the diagonal versus the main diagonal. 0 corresponds to the main diagonal, a negative/positive value to diagonals below/above the main diagonal.
  • lower (boolean, optional, default=1) – Refer to the lower triangular matrix if lower=true, refer to the upper otherwise. Only relevant when offset=0
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_gelqf(A=None, out=None, name=None, **kwargs)

LQ factorization for general matrix. Input is a tensor A of dimension n >= 2.

If n=2, we compute the LQ factorization (LAPACK gelqf, followed by orglq). A must have shape (x, y) with x <= y, and must have full rank =x. The LQ factorization consists of L with shape (x, x) and Q with shape (x, y), so that:

A = L * Q

Here, L is lower triangular (upper triangle equal to zero) with nonzero diagonal, and Q is row-orthonormal, meaning that

Q * QT

is equal to the identity matrix of shape (x, x).

If n>2, gelqf is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Examples:

// Single LQ factorization
A = [[1., 2., 3.], [4., 5., 6.]]
Q, L = gelqf(A)
Q = [[-0.26726124, -0.53452248, -0.80178373],
     [0.87287156, 0.21821789, -0.43643578]]
L = [[-3.74165739, 0.],
     [-8.55235974, 1.96396101]]

// Batch LQ factorization
A = [[[1., 2., 3.], [4., 5., 6.]],
     [[7., 8., 9.], [10., 11., 12.]]]
Q, L = gelqf(A)
Q = [[[-0.26726124, -0.53452248, -0.80178373],
      [0.87287156, 0.21821789, -0.43643578]],
     [[-0.50257071, -0.57436653, -0.64616234],
      [0.7620735, 0.05862104, -0.64483142]]]
L = [[[-3.74165739, 0.],
      [-8.55235974, 1.96396101]],
     [[-13.92838828, 0.],
      [-19.09768702, 0.52758934]]]

Defined in src/operator/tensor/la_op.cc:L798

Parameters:
  • A (NDArray) – Tensor of input matrices to be factorized
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_gemm(A=None, B=None, C=None, transpose_a=_Null, transpose_b=_Null, alpha=_Null, beta=_Null, axis=_Null, out=None, name=None, **kwargs)

Performs general matrix multiplication and accumulation. Input are tensors A, B, C, each of dimension n >= 2 and having the same shape on the leading n-2 dimensions.

If n=2, the BLAS3 function gemm is performed:

out = alpha * op(A) * op(B) + beta * C

Here, alpha and beta are scalar parameters, and op() is either the identity or matrix transposition (depending on transpose_a, transpose_b).

If n>2, gemm is performed separately for a batch of matrices. The column indices of the matrices are given by the last dimensions of the tensors, the row indices by the axis specified with the axis parameter. By default, the trailing two dimensions will be used for matrix encoding.

For a non-default axis parameter, the operation performed is equivalent to a series of swapaxes/gemm/swapaxes calls. For example let A, B, C be 5 dimensional tensors. Then gemm(A, B, C, axis=1) is equivalent to the following without the overhead of the additional swapaxis operations:

A1 = swapaxes(A, dim1=1, dim2=3)
B1 = swapaxes(B, dim1=1, dim2=3)
C = swapaxes(C, dim1=1, dim2=3)
C = gemm(A1, B1, C)
C = swapaxis(C, dim1=1, dim2=3)

When the input data is of type float32 and the environment variables MXNET_CUDA_ALLOW_TENSOR_CORE and MXNET_CUDA_TENSOR_OP_MATH_ALLOW_CONVERSION are set to 1, this operator will try to use pseudo-float16 precision (float32 math with float16 I/O) precision in order to use Tensor Cores on suitable NVIDIA GPUs. This can sometimes give significant speedups.

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix multiply-add
A = [[1.0, 1.0], [1.0, 1.0]]
B = [[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]]
C = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]
gemm(A, B, C, transpose_b=True, alpha=2.0, beta=10.0)
        = [[14.0, 14.0, 14.0], [14.0, 14.0, 14.0]]

// Batch matrix multiply-add
A = [[[1.0, 1.0]], [[0.1, 0.1]]]
B = [[[1.0, 1.0]], [[0.1, 0.1]]]
C = [[[10.0]], [[0.01]]]
gemm(A, B, C, transpose_b=True, alpha=2.0 , beta=10.0)
        = [[[104.0]], [[0.14]]]

Defined in src/operator/tensor/la_op.cc:L89

Parameters:
  • A (NDArray) – Tensor of input matrices
  • B (NDArray) – Tensor of input matrices
  • C (NDArray) – Tensor of input matrices
  • transpose_a (boolean, optional, default=0) – Multiply with transposed of first input (A).
  • transpose_b (boolean, optional, default=0) – Multiply with transposed of second input (B).
  • alpha (double, optional, default=1) – Scalar factor multiplied with A*B.
  • beta (double, optional, default=1) – Scalar factor multiplied with C.
  • axis (int, optional, default='-2') – Axis corresponding to the matrix rows.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_gemm2(A=None, B=None, transpose_a=_Null, transpose_b=_Null, alpha=_Null, axis=_Null, out=None, name=None, **kwargs)

Performs general matrix multiplication. Input are tensors A, B, each of dimension n >= 2 and having the same shape on the leading n-2 dimensions.

If n=2, the BLAS3 function gemm is performed:

out = alpha * op(A) * op(B)

Here alpha is a scalar parameter and op() is either the identity or the matrix transposition (depending on transpose_a, transpose_b).

If n>2, gemm is performed separately for a batch of matrices. The column indices of the matrices are given by the last dimensions of the tensors, the row indices by the axis specified with the axis parameter. By default, the trailing two dimensions will be used for matrix encoding.

For a non-default axis parameter, the operation performed is equivalent to a series of swapaxes/gemm/swapaxes calls. For example let A, B be 5 dimensional tensors. Then gemm(A, B, axis=1) is equivalent to the following without the overhead of the additional swapaxis operations:

A1 = swapaxes(A, dim1=1, dim2=3)
B1 = swapaxes(B, dim1=1, dim2=3)
C = gemm2(A1, B1)
C = swapaxis(C, dim1=1, dim2=3)

When the input data is of type float32 and the environment variables MXNET_CUDA_ALLOW_TENSOR_CORE and MXNET_CUDA_TENSOR_OP_MATH_ALLOW_CONVERSION are set to 1, this operator will try to use pseudo-float16 precision (float32 math with float16 I/O) precision in order to use Tensor Cores on suitable NVIDIA GPUs. This can sometimes give significant speedups.

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix multiply
A = [[1.0, 1.0], [1.0, 1.0]]
B = [[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]]
gemm2(A, B, transpose_b=True, alpha=2.0)
         = [[4.0, 4.0, 4.0], [4.0, 4.0, 4.0]]

// Batch matrix multiply
A = [[[1.0, 1.0]], [[0.1, 0.1]]]
B = [[[1.0, 1.0]], [[0.1, 0.1]]]
gemm2(A, B, transpose_b=True, alpha=2.0)
        = [[[4.0]], [[0.04 ]]]

Defined in src/operator/tensor/la_op.cc:L163

Parameters:
  • A (NDArray) – Tensor of input matrices
  • B (NDArray) – Tensor of input matrices
  • transpose_a (boolean, optional, default=0) – Multiply with transposed of first input (A).
  • transpose_b (boolean, optional, default=0) – Multiply with transposed of second input (B).
  • alpha (double, optional, default=1) – Scalar factor multiplied with A*B.
  • axis (int, optional, default='-2') – Axis corresponding to the matrix row indices.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_inverse(A=None, out=None, name=None, **kwargs)

Compute the inverse of a matrix. Input is a tensor A of dimension n >= 2.

If n=2, A is a square matrix. We compute:

out = A-1

If n>2, inverse is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix inversion
A = [[1., 4.], [2., 3.]]
inverse(A) = [[-0.6, 0.8], [0.4, -0.2]]

// Batch matrix inversion
A = [[[1., 4.], [2., 3.]],
     [[1., 3.], [2., 4.]]]
inverse(A) = [[[-0.6, 0.8], [0.4, -0.2]],
              [[-2., 1.5], [1., -0.5]]]

Defined in src/operator/tensor/la_op.cc:L917

Parameters:
  • A (NDArray) – Tensor of square matrix
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_makediag(A=None, offset=_Null, out=None, name=None, **kwargs)

Constructs a square matrix with the input as diagonal. Input is a tensor A of dimension n >= 1.

If n=1, then A represents the diagonal entries of a single square matrix. This matrix will be returned as a 2-dimensional tensor. If n>1, then A represents a batch of diagonals of square matrices. The batch of diagonal matrices will be returned as an n+1-dimensional tensor.

Note

The operator supports float32 and float64 data types only.

Examples:

// Single diagonal matrix construction
A = [1.0, 2.0]

makediag(A)    = [[1.0, 0.0],
                  [0.0, 2.0]]

makediag(A, 1) = [[0.0, 1.0, 0.0],
                  [0.0, 0.0, 2.0],
                  [0.0, 0.0, 0.0]]

// Batch diagonal matrix construction
A = [[1.0, 2.0],
     [3.0, 4.0]]

makediag(A) = [[[1.0, 0.0],
                [0.0, 2.0]],
               [[3.0, 0.0],
                [0.0, 4.0]]]

Defined in src/operator/tensor/la_op.cc:L547

Parameters:
  • A (NDArray) – Tensor of diagonal entries
  • offset (int, optional, default='0') – Offset of the diagonal versus the main diagonal. 0 corresponds to the main diagonal, a negative/positive value to diagonals below/above the main diagonal.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_maketrian(A=None, offset=_Null, lower=_Null, out=None, name=None, **kwargs)

Constructs a square matrix with the input representing a specific triangular sub-matrix. This is basically the inverse of linalg.extracttrian. Input is a tensor A of dimension n >= 1.

If n=1, then A represents the entries of a triangular matrix which is lower triangular if offset<0 or offset=0, lower=true. The resulting matrix is derived by first constructing the square matrix with the entries outside the triangle set to zero and then adding offset-times an additional diagonal with zero entries to the square matrix.

If n>1, then A represents a batch of triangular sub-matrices. The batch of corresponding square matrices is returned as an n+1-dimensional tensor.

Note

The operator supports float32 and float64 data types only.

Examples:

// Single  matrix construction
A = [1.0, 2.0, 3.0]

maketrian(A)              = [[1.0, 0.0],
                             [2.0, 3.0]]

maketrian(A, lower=false) = [[1.0, 2.0],
                             [0.0, 3.0]]

maketrian(A, offset=1)    = [[0.0, 1.0, 2.0],
                             [0.0, 0.0, 3.0],
                             [0.0, 0.0, 0.0]]
maketrian(A, offset=-1)   = [[0.0, 0.0, 0.0],
                             [1.0, 0.0, 0.0],
                             [2.0, 3.0, 0.0]]

// Batch matrix construction
A = [[1.0, 2.0, 3.0],
     [4.0, 5.0, 6.0]]

maketrian(A)           = [[[1.0, 0.0],
                           [2.0, 3.0]],
                          [[4.0, 0.0],
                           [5.0, 6.0]]]

maketrian(A, offset=1) = [[[0.0, 1.0, 2.0],
                           [0.0, 0.0, 3.0],
                           [0.0, 0.0, 0.0]],
                          [[0.0, 4.0, 5.0],
                           [0.0, 0.0, 6.0],
                           [0.0, 0.0, 0.0]]]

Defined in src/operator/tensor/la_op.cc:L673

Parameters:
  • A (NDArray) – Tensor of triangular matrices stored as vectors
  • offset (int, optional, default='0') – Offset of the diagonal versus the main diagonal. 0 corresponds to the main diagonal, a negative/positive value to diagonals below/above the main diagonal.
  • lower (boolean, optional, default=1) – Refer to the lower triangular matrix if lower=true, refer to the upper otherwise. Only relevant when offset=0
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_potrf(A=None, out=None, name=None, **kwargs)

Performs Cholesky factorization of a symmetric positive-definite matrix. Input is a tensor A of dimension n >= 2.

If n=2, the Cholesky factor B of the symmetric, positive definite matrix A is computed. B is triangular (entries of upper or lower triangle are all zero), has positive diagonal entries, and:

A = B * BT if lower = true A = BT * B if lower = false

If n>2, potrf is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix factorization
A = [[4.0, 1.0], [1.0, 4.25]]
potrf(A) = [[2.0, 0], [0.5, 2.0]]

// Batch matrix factorization
A = [[[4.0, 1.0], [1.0, 4.25]], [[16.0, 4.0], [4.0, 17.0]]]
potrf(A) = [[[2.0, 0], [0.5, 2.0]], [[4.0, 0], [1.0, 4.0]]]

Defined in src/operator/tensor/la_op.cc:L214

Parameters:
  • A (NDArray) – Tensor of input matrices to be decomposed
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_potri(A=None, out=None, name=None, **kwargs)

Performs matrix inversion from a Cholesky factorization. Input is a tensor A of dimension n >= 2.

If n=2, A is a triangular matrix (entries of upper or lower triangle are all zero) with positive diagonal. We compute:

out = A-T * A-1 if lower = true out = A-1 * A-T if lower = false

In other words, if A is the Cholesky factor of a symmetric positive definite matrix B (obtained by potrf), then

out = B-1

If n>2, potri is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Note

Use this operator only if you are certain you need the inverse of B, and cannot use the Cholesky factor A (potrf), together with backsubstitution (trsm). The latter is numerically much safer, and also cheaper.

Examples:

// Single matrix inverse
A = [[2.0, 0], [0.5, 2.0]]
potri(A) = [[0.26563, -0.0625], [-0.0625, 0.25]]

// Batch matrix inverse
A = [[[2.0, 0], [0.5, 2.0]], [[4.0, 0], [1.0, 4.0]]]
potri(A) = [[[0.26563, -0.0625], [-0.0625, 0.25]],
            [[0.06641, -0.01562], [-0.01562, 0,0625]]]

Defined in src/operator/tensor/la_op.cc:L275

Parameters:
  • A (NDArray) – Tensor of lower triangular matrices
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_sumlogdiag(A=None, out=None, name=None, **kwargs)

Computes the sum of the logarithms of the diagonal elements of a square matrix. Input is a tensor A of dimension n >= 2.

If n=2, A must be square with positive diagonal entries. We sum the natural logarithms of the diagonal elements, the result has shape (1,).

If n>2, sumlogdiag is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix reduction
A = [[1.0, 1.0], [1.0, 7.0]]
sumlogdiag(A) = [1.9459]

// Batch matrix reduction
A = [[[1.0, 1.0], [1.0, 7.0]], [[3.0, 0], [0, 17.0]]]
sumlogdiag(A) = [1.9459, 3.9318]

Defined in src/operator/tensor/la_op.cc:L445

Parameters:
  • A (NDArray) – Tensor of square matrices
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_syrk(A=None, transpose=_Null, alpha=_Null, out=None, name=None, **kwargs)

Multiplication of matrix with its transpose. Input is a tensor A of dimension n >= 2.

If n=2, the operator performs the BLAS3 function syrk:

out = alpha * A * AT

if transpose=False, or

out = alpha * AT * A

if transpose=True.

If n>2, syrk is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix multiply
A = [[1., 2., 3.], [4., 5., 6.]]
syrk(A, alpha=1., transpose=False)
         = [[14., 32.],
            [32., 77.]]
syrk(A, alpha=1., transpose=True)
         = [[17., 22., 27.],
            [22., 29., 36.],
            [27., 36., 45.]]

// Batch matrix multiply
A = [[[1., 1.]], [[0.1, 0.1]]]
syrk(A, alpha=2., transpose=False) = [[[4.]], [[0.04]]]

Defined in src/operator/tensor/la_op.cc:L730

Parameters:
  • A (NDArray) – Tensor of input matrices
  • transpose (boolean, optional, default=0) – Use transpose of input matrix.
  • alpha (double, optional, default=1) – Scalar factor to be applied to the result.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_trmm(A=None, B=None, transpose=_Null, rightside=_Null, lower=_Null, alpha=_Null, out=None, name=None, **kwargs)

Performs multiplication with a lower triangular matrix. Input are tensors A, B, each of dimension n >= 2 and having the same shape on the leading n-2 dimensions.

If n=2, A must be triangular. The operator performs the BLAS3 function trmm:

out = alpha * op(A) * B

if rightside=False, or

out = alpha * B * op(A)

if rightside=True. Here, alpha is a scalar parameter, and op() is either the identity or the matrix transposition (depending on transpose).

If n>2, trmm is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Examples:

// Single triangular matrix multiply
A = [[1.0, 0], [1.0, 1.0]]
B = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]
trmm(A, B, alpha=2.0) = [[2.0, 2.0, 2.0], [4.0, 4.0, 4.0]]

// Batch triangular matrix multiply
A = [[[1.0, 0], [1.0, 1.0]], [[1.0, 0], [1.0, 1.0]]]
B = [[[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]], [[0.5, 0.5, 0.5], [0.5, 0.5, 0.5]]]
trmm(A, B, alpha=2.0) = [[[2.0, 2.0, 2.0], [4.0, 4.0, 4.0]],
                         [[1.0, 1.0, 1.0], [2.0, 2.0, 2.0]]]

Defined in src/operator/tensor/la_op.cc:L333

Parameters:
  • A (NDArray) – Tensor of lower triangular matrices
  • B (NDArray) – Tensor of matrices
  • transpose (boolean, optional, default=0) – Use transposed of the triangular matrix
  • rightside (boolean, optional, default=0) – Multiply triangular matrix from the right to non-triangular one.
  • lower (boolean, optional, default=1) – True if the triangular matrix is lower triangular, false if it is upper triangular.
  • alpha (double, optional, default=1) – Scalar factor to be applied to the result.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.linalg_trsm(A=None, B=None, transpose=_Null, rightside=_Null, lower=_Null, alpha=_Null, out=None, name=None, **kwargs)

Solves matrix equation involving a lower triangular matrix. Input are tensors A, B, each of dimension n >= 2 and having the same shape on the leading n-2 dimensions.

If n=2, A must be triangular. The operator performs the BLAS3 function trsm, solving for out in:

op(A) * out = alpha * B

if rightside=False, or

out * op(A) = alpha * B

if rightside=True. Here, alpha is a scalar parameter, and op() is either the identity or the matrix transposition (depending on transpose).

If n>2, trsm is performed separately on the trailing two dimensions for all inputs (batch mode).

Note

The operator supports float32 and float64 data types only.

Examples:

// Single matrix solve
A = [[1.0, 0], [1.0, 1.0]]
B = [[2.0, 2.0, 2.0], [4.0, 4.0, 4.0]]
trsm(A, B, alpha=0.5) = [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]

// Batch matrix solve
A = [[[1.0, 0], [1.0, 1.0]], [[1.0, 0], [1.0, 1.0]]]
B = [[[2.0, 2.0, 2.0], [4.0, 4.0, 4.0]],
     [[4.0, 4.0, 4.0], [8.0, 8.0, 8.0]]]
trsm(A, B, alpha=0.5) = [[[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
                         [[2.0, 2.0, 2.0], [2.0, 2.0, 2.0]]]

Defined in src/operator/tensor/la_op.cc:L396

Parameters:
  • A (NDArray) – Tensor of lower triangular matrices
  • B (NDArray) – Tensor of matrices
  • transpose (boolean, optional, default=0) – Use transposed of the triangular matrix
  • rightside (boolean, optional, default=0) – Multiply triangular matrix from the right to non-triangular one.
  • lower (boolean, optional, default=1) – True if the triangular matrix is lower triangular, false if it is upper triangular.
  • alpha (double, optional, default=1) – Scalar factor to be applied to the result.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.log(data=None, out=None, name=None, **kwargs)

Returns element-wise Natural logarithmic value of the input.

The natural logarithm is logarithm in base e, so that log(exp(x)) = x

The storage type of log output is always dense

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L1057

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.log10(data=None, out=None, name=None, **kwargs)

Returns element-wise Base-10 logarithmic value of the input.

10**log10(x) = x

The storage type of log10 output is always dense

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L1074

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.log1p(data=None, out=None, name=None, **kwargs)

Returns element-wise log(1 + x) value of the input.

This function is more accurate than log(1 + x) for small x so that \(1+x\approx 1\)

The storage type of log1p output depends upon the input storage type:

  • log1p(default) = default
  • log1p(row_sparse) = row_sparse
  • log1p(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L1171

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.log2(data=None, out=None, name=None, **kwargs)

Returns element-wise Base-2 logarithmic value of the input.

2**log2(x) = x

The storage type of log2 output is always dense

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L1086

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.log_softmax(data=None, axis=_Null, temperature=_Null, dtype=_Null, out=None, name=None, **kwargs)

Computes the log softmax of the input. This is equivalent to computing softmax followed by log.

Examples:

>>> x = mx.nd.array([1, 2, .1])
>>> mx.nd.log_softmax(x).asnumpy()
array([-1.41702998, -0.41702995, -2.31702995], dtype=float32)

>>> x = mx.nd.array( [[1, 2, .1],[.1, 2, 1]] )
>>> mx.nd.log_softmax(x, axis=0).asnumpy()
array([[-0.34115392, -0.69314718, -1.24115396],
       [-1.24115396, -0.69314718, -0.34115392]], dtype=float32)
Parameters:
  • data (NDArray) – The input array.
  • axis (int, optional, default='-1') – The axis along which to compute softmax.
  • temperature (double or None, optional, default=None) – Temperature parameter in softmax
  • dtype ({None, 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to the same as input’s dtype if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.logical_not(data=None, out=None, name=None, **kwargs)

Returns the result of logical NOT (!) function

Example

logical_not([-2., 0., 1.]) = [0., 1., 0.]

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.make_loss(data=None, out=None, name=None, **kwargs)

Make your own loss function in network construction.

This operator accepts a customized loss function symbol as a terminal loss and the symbol should be an operator with no backward dependency. The output of this function is the gradient of loss with respect to the input data.

For example, if you are a making a cross entropy loss function. Assume out is the predicted output and label is the true label, then the cross entropy can be defined as:

cross_entropy = label * log(out) + (1 - label) * log(1 - out)
loss = make_loss(cross_entropy)

We will need to use make_loss when we are creating our own loss function or we want to combine multiple loss functions. Also we may want to stop some variables’ gradients from backpropagation. See more detail in BlockGrad or stop_gradient.

The storage type of make_loss output depends upon the input storage type:

  • make_loss(default) = default
  • make_loss(row_sparse) = row_sparse

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L332

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.max(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the max of array elements over given axes.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L191

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.max_axis(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the max of array elements over given axes.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L191

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.mean(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the mean of array elements over given axes.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L132

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.min(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the min of array elements over given axes.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L205

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.min_axis(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the min of array elements over given axes.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L205

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.moments(data=None, axes=_Null, keepdims=_Null, out=None, name=None, **kwargs)

Calculate the mean and variance of data.

The mean and variance are calculated by aggregating the contents of data across axes. If x is 1-D and axes = [0] this is just the mean and variance of a vector.

Example

x = [[1, 2, 3], [4, 5, 6]] mean, var = moments(data=x, axes=[0]) mean = [2.5, 3.5, 4.5] var = [2.25, 2.25, 2.25] mean, var = moments(data=x, axes=[1]) mean = [2.0, 5.0] var = [0.66666667, 0.66666667] mean, var = moments(data=x, axis=[0, 1]) mean = [3.5] var = [2.9166667]

Defined in src/operator/nn/moments.cc:L54

Parameters:
  • data (NDArray) – Input ndarray
  • axes (Shape or None, optional, default=None) – Array of ints. Axes along which to compute mean and variance.
  • keepdims (boolean, optional, default=0) – produce moments with the same dimensionality as the input.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.mp_nag_mom_update(weight=None, grad=None, mom=None, weight32=None, lr=_Null, momentum=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, out=None, name=None, **kwargs)

Update function for multi-precision Nesterov Accelerated Gradient( NAG) optimizer.

Defined in src/operator/optimizer_op.cc:L743

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • mom (NDArray) – Momentum
  • weight32 (NDArray) – Weight32
  • lr (float, required) – Learning rate
  • momentum (float, optional, default=0) – The decay rate of momentum estimates at each epoch.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.mp_sgd_mom_update(weight=None, grad=None, mom=None, weight32=None, lr=_Null, momentum=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, lazy_update=_Null, out=None, name=None, **kwargs)

Updater function for multi-precision sgd optimizer

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • mom (NDArray) – Momentum
  • weight32 (NDArray) – Weight32
  • lr (float, required) – Learning rate
  • momentum (float, optional, default=0) – The decay rate of momentum estimates at each epoch.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • lazy_update (boolean, optional, default=1) – If true, lazy updates are applied if gradient’s stype is row_sparse and both weight and momentum have the same stype
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.mp_sgd_update(weight=None, grad=None, weight32=None, lr=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, lazy_update=_Null, out=None, name=None, **kwargs)

Updater function for multi-precision sgd optimizer

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – gradient
  • weight32 (NDArray) – Weight32
  • lr (float, required) – Learning rate
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • lazy_update (boolean, optional, default=1) – If true, lazy updates are applied if gradient’s stype is row_sparse.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.multi_all_finite(*data, **kwargs)

Check if all the float numbers in all the arrays are finite (used for AMP)

Defined in src/operator/contrib/all_finite.cc:L133

Parameters:
  • data (NDArray[]) – Arrays
  • num_arrays (int, optional, default='1') – Number of arrays.
  • init_output (boolean, optional, default=1) – Initialize output to 1.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.multi_mp_sgd_mom_update(*data, **kwargs)

Momentum update function for multi-precision Stochastic Gradient Descent (SGD) optimizer.

Momentum update has better convergence rates on neural networks. Mathematically it looks like below:

\[\begin{split}v_1 = \alpha * \nabla J(W_0)\\ v_t = \gamma v_{t-1} - \alpha * \nabla J(W_{t-1})\\ W_t = W_{t-1} + v_t\end{split}\]

It updates the weights using:

v = momentum * v - learning_rate * gradient
weight += v

Where the parameter momentum is the decay rate of momentum estimates at each epoch.

Defined in src/operator/optimizer_op.cc:L470

Parameters:
  • data (NDArray[]) – Weights
  • lrs (tuple of , required) – Learning rates.
  • wds (tuple of , required) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • momentum (float, optional, default=0) – The decay rate of momentum estimates at each epoch.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • num_weights (int, optional, default='1') – Number of updated weights.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.multi_mp_sgd_update(*data, **kwargs)

Update function for multi-precision Stochastic Gradient Descent (SDG) optimizer.

It updates the weights using:

weight = weight - learning_rate * (gradient + wd * weight)

Defined in src/operator/optimizer_op.cc:L415

Parameters:
  • data (NDArray[]) – Weights
  • lrs (tuple of , required) – Learning rates.
  • wds (tuple of , required) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • num_weights (int, optional, default='1') – Number of updated weights.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.multi_sgd_mom_update(*data, **kwargs)

Momentum update function for Stochastic Gradient Descent (SGD) optimizer.

Momentum update has better convergence rates on neural networks. Mathematically it looks like below:

\[\begin{split}v_1 = \alpha * \nabla J(W_0)\\ v_t = \gamma v_{t-1} - \alpha * \nabla J(W_{t-1})\\ W_t = W_{t-1} + v_t\end{split}\]

It updates the weights using:

v = momentum * v - learning_rate * gradient
weight += v

Where the parameter momentum is the decay rate of momentum estimates at each epoch.

Defined in src/operator/optimizer_op.cc:L372

Parameters:
  • data (NDArray[]) – Weights, gradients and momentum
  • lrs (tuple of , required) – Learning rates.
  • wds (tuple of , required) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • momentum (float, optional, default=0) – The decay rate of momentum estimates at each epoch.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • num_weights (int, optional, default='1') – Number of updated weights.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.multi_sgd_update(*data, **kwargs)

Update function for Stochastic Gradient Descent (SDG) optimizer.

It updates the weights using:

weight = weight - learning_rate * (gradient + wd * weight)

Defined in src/operator/optimizer_op.cc:L327

Parameters:
  • data (NDArray[]) – Weights
  • lrs (tuple of , required) – Learning rates.
  • wds (tuple of , required) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • num_weights (int, optional, default='1') – Number of updated weights.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.nag_mom_update(weight=None, grad=None, mom=None, lr=_Null, momentum=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, out=None, name=None, **kwargs)

Update function for Nesterov Accelerated Gradient( NAG) optimizer. It updates the weights using the following formula,

\[\begin{split}v_t = \gamma v_{t-1} + \eta * \nabla J(W_{t-1} - \gamma v_{t-1})\\ W_t = W_{t-1} - v_t\end{split}\]

Where \(\eta\) is the learning rate of the optimizer \(\gamma\) is the decay rate of the momentum estimate \(\v_t\) is the update vector at time step t \(\W_t\) is the weight vector at time step t

Defined in src/operator/optimizer_op.cc:L724

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • mom (NDArray) – Momentum
  • lr (float, required) – Learning rate
  • momentum (float, optional, default=0) – The decay rate of momentum estimates at each epoch.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.nanprod(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the product of array elements over given axes treating Not a Numbers (NaN) as one.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L177

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.nansum(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the sum of array elements over given axes treating Not a Numbers (NaN) as zero.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L162

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.negative(data=None, out=None, name=None, **kwargs)

Numerical negative of the argument, element-wise.

The storage type of negative output depends upon the input storage type:

  • negative(default) = default
  • negative(row_sparse) = row_sparse
  • negative(csr) = csr
Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.norm(data=None, ord=_Null, axis=_Null, out_dtype=_Null, keepdims=_Null, out=None, name=None, **kwargs)

Computes the norm on an NDArray.

This operator computes the norm on an NDArray with the specified axis, depending on the value of the ord parameter. By default, it computes the L2 norm on the entire array. Currently only ord=2 supports sparse ndarrays.

Examples:

x = [[[1, 2],
      [3, 4]],
     [[2, 2],
      [5, 6]]]

norm(x, ord=2, axis=1) = [[3.1622777 4.472136 ]
                          [5.3851647 6.3245554]]

norm(x, ord=1, axis=1) = [[4., 6.],
                          [7., 8.]]

rsp = x.cast_storage('row_sparse')

norm(rsp) = [5.47722578]

csr = x.cast_storage('csr')

norm(csr) = [5.47722578]

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L350

Parameters:
  • data (NDArray) – The input
  • ord (int, optional, default='2') – Order of the norm. Currently ord=1 and ord=2 is supported.
  • axis (Shape or None, optional, default=None) –
    The axis or axes along which to perform the reduction.
    The default, axis=(), will compute over all elements into a scalar array with shape (1,). If axis is int, a reduction is performed on a particular axis. If axis is a 2-tuple, it specifies the axes that hold 2-D matrices, and the matrix norms of these matrices are computed.
  • out_dtype ({None, 'float16', 'float32', 'float64', 'int32', 'int64', 'int8'},optional, default='None') – The data type of the output.
  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axis is left in the result as dimension with size one.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.normal(loc=_Null, scale=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a normal (Gaussian) distribution.

Note

The existing alias normal is deprecated.

Samples are distributed according to a normal distribution parametrized by loc (mean) and scale (standard deviation).

Example:

normal(loc=0, scale=1, shape=(2,2)) = [[ 1.89171135, -1.16881478],
                                       [-1.23474145,  1.55807114]]

Defined in src/operator/random/sample_op.cc:L113

Parameters:
  • loc (float, optional, default=0) – Mean of the distribution.
  • scale (float, optional, default=1) – Standard deviation of the distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.one_hot(indices=None, depth=_Null, on_value=_Null, off_value=_Null, dtype=_Null, out=None, name=None, **kwargs)

Returns a one-hot array.

The locations represented by indices take value on_value, while all other locations take value off_value.

one_hot operation with indices of shape (i0, i1) and depth of d would result in an output array of shape (i0, i1, d) with:

output[i,j,:] = off_value
output[i,j,indices[i,j]] = on_value

Examples:

one_hot([1,0,2,0], 3) = [[ 0.  1.  0.]
                         [ 1.  0.  0.]
                         [ 0.  0.  1.]
                         [ 1.  0.  0.]]

one_hot([1,0,2,0], 3, on_value=8, off_value=1,
        dtype='int32') = [[1 8 1]
                          [8 1 1]
                          [1 1 8]
                          [8 1 1]]

one_hot([[1,0],[1,0],[2,0]], 3) = [[[ 0.  1.  0.]
                                    [ 1.  0.  0.]]

                                   [[ 0.  1.  0.]
                                    [ 1.  0.  0.]]

                                   [[ 0.  0.  1.]
                                    [ 1.  0.  0.]]]

Defined in src/operator/tensor/indexing_op.cc:L799

Parameters:
  • indices (NDArray) – array of locations where to set on_value
  • depth (int, required) – Depth of the one hot dimension.
  • on_value (double, optional, default=1) – The value assigned to the locations represented by indices.
  • off_value (double, optional, default=0) – The value assigned to the locations not represented by indices.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8'},optional, default='float32') – DType of the output
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ones_like(data=None, out=None, name=None, **kwargs)

Return an array of ones with the same shape and type as the input array.

Examples:

x = [[ 0.,  0.,  0.],
     [ 0.,  0.,  0.]]

ones_like(x) = [[ 1.,  1.,  1.],
                [ 1.,  1.,  1.]]
Parameters:
  • data (NDArray) – The input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.pad(data=None, mode=_Null, pad_width=_Null, constant_value=_Null, out=None, name=None, **kwargs)

Pads an input array with a constant or edge values of the array.

Note

Pad is deprecated. Use pad instead.

Note

Current implementation only supports 4D and 5D input arrays with padding applied only on axes 1, 2 and 3. Expects axes 4 and 5 in pad_width to be zero.

This operation pads an input array with either a constant_value or edge values along each axis of the input array. The amount of padding is specified by pad_width.

pad_width is a tuple of integer padding widths for each axis of the format (before_1, after_1, ... , before_N, after_N). The pad_width should be of length 2*N where N is the number of dimensions of the array.

For dimension N of the input array, before_N and after_N indicates how many values to add before and after the elements of the array along dimension N. The widths of the higher two dimensions before_1, after_1, before_2, after_2 must be 0.

Example:

x = [[[[  1.   2.   3.]
       [  4.   5.   6.]]

      [[  7.   8.   9.]
       [ 10.  11.  12.]]]


     [[[ 11.  12.  13.]
       [ 14.  15.  16.]]

      [[ 17.  18.  19.]
       [ 20.  21.  22.]]]]

pad(x,mode="edge", pad_width=(0,0,0,0,1,1,1,1)) =

      [[[[  1.   1.   2.   3.   3.]
         [  1.   1.   2.   3.   3.]
         [  4.   4.   5.   6.   6.]
         [  4.   4.   5.   6.   6.]]

        [[  7.   7.   8.   9.   9.]
         [  7.   7.   8.   9.   9.]
         [ 10.  10.  11.  12.  12.]
         [ 10.  10.  11.  12.  12.]]]


       [[[ 11.  11.  12.  13.  13.]
         [ 11.  11.  12.  13.  13.]
         [ 14.  14.  15.  16.  16.]
         [ 14.  14.  15.  16.  16.]]

        [[ 17.  17.  18.  19.  19.]
         [ 17.  17.  18.  19.  19.]
         [ 20.  20.  21.  22.  22.]
         [ 20.  20.  21.  22.  22.]]]]

pad(x, mode="constant", constant_value=0, pad_width=(0,0,0,0,1,1,1,1)) =

      [[[[  0.   0.   0.   0.   0.]
         [  0.   1.   2.   3.   0.]
         [  0.   4.   5.   6.   0.]
         [  0.   0.   0.   0.   0.]]

        [[  0.   0.   0.   0.   0.]
         [  0.   7.   8.   9.   0.]
         [  0.  10.  11.  12.   0.]
         [  0.   0.   0.   0.   0.]]]


       [[[  0.   0.   0.   0.   0.]
         [  0.  11.  12.  13.   0.]
         [  0.  14.  15.  16.   0.]
         [  0.   0.   0.   0.   0.]]

        [[  0.   0.   0.   0.   0.]
         [  0.  17.  18.  19.   0.]
         [  0.  20.  21.  22.   0.]
         [  0.   0.   0.   0.   0.]]]]

Defined in src/operator/pad.cc:L766

Parameters:
  • data (NDArray) – An n-dimensional input array.
  • mode ({'constant', 'edge', 'reflect'}, required) – Padding type to use. “constant” pads with constant_value “edge” pads using the edge values of the input array “reflect” pads by reflecting values with respect to the edges.
  • pad_width (Shape(tuple), required) – Widths of the padding regions applied to the edges of each axis. It is a tuple of integer padding widths for each axis of the format (before_1, after_1, ... , before_N, after_N). It should be of length 2*N where N is the number of dimensions of the array.This is equivalent to pad_width in numpy.pad, but flattened.
  • constant_value (double, optional, default=0) – The value used for padding when mode is “constant”.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.pick(data=None, index=None, axis=_Null, keepdims=_Null, mode=_Null, out=None, name=None, **kwargs)

Picks elements from an input array according to the input indices along the given axis.

Given an input array of shape (d0, d1) and indices of shape (i0,), the result will be an output array of shape (i0,) with:

output[i] = input[i, indices[i]]

By default, if any index mentioned is too large, it is replaced by the index that addresses the last element along an axis (the clip mode).

This function supports n-dimensional input and (n-1)-dimensional indices arrays.

Examples:

x = [[ 1.,  2.],
     [ 3.,  4.],
     [ 5.,  6.]]

// picks elements with specified indices along axis 0
pick(x, y=[0,1], 0) = [ 1.,  4.]

// picks elements with specified indices along axis 1
pick(x, y=[0,1,0], 1) = [ 1.,  4.,  5.]

y = [[ 1.],
     [ 0.],
     [ 2.]]

// picks elements with specified indices along axis 1 using 'wrap' mode
// to place indicies that would normally be out of bounds
pick(x, y=[2,-1,-2], 1, mode='wrap') = [ 1.,  4.,  5.]

y = [[ 1.],
     [ 0.],
     [ 2.]]

// picks elements with specified indices along axis 1 and dims are maintained
pick(x,y, 1, keepdims=True) = [[ 2.],
                               [ 3.],
                               [ 6.]]

Defined in src/operator/tensor/broadcast_reduce_op_index.cc:L154

Parameters:
  • data (NDArray) – The input array
  • index (NDArray) – The index array
  • axis (int or None, optional, default='-1') – int or None. The axis to picking the elements. Negative values means indexing from right to left. If is None, the elements in the index w.r.t the flattened input will be picked.
  • keepdims (boolean, optional, default=0) – If true, the axis where we pick the elements is left in the result as dimension with size one.
  • mode ({'clip', 'wrap'},optional, default='clip') – Specify how out-of-bound indices behave. Default is “clip”. “clip” means clip to the range. So, if all indices mentioned are too large, they are replaced by the index that addresses the last element along an axis. “wrap” means to wrap around.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.prod(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the product of array elements over given axes.

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L147

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.radians(data=None, out=None, name=None, **kwargs)

Converts each element of the input array from degrees to radians.

\[radians([0, 90, 180, 270, 360]) = [0, \pi/2, \pi, 3\pi/2, 2\pi]\]

The storage type of radians output depends upon the input storage type:

  • radians(default) = default
  • radians(row_sparse) = row_sparse
  • radians(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L238

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_exponential(lam=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from an exponential distribution.

Samples are distributed according to an exponential distribution parametrized by lambda (rate).

Example:

exponential(lam=4, shape=(2,2)) = [[ 0.0097189 ,  0.08999364],
                                   [ 0.04146638,  0.31715935]]

Defined in src/operator/random/sample_op.cc:L137

Parameters:
  • lam (float, optional, default=1) – Lambda parameter (rate) of the exponential distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_gamma(alpha=_Null, beta=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a gamma distribution.

Samples are distributed according to a gamma distribution parametrized by alpha (shape) and beta (scale).

Example:

gamma(alpha=9, beta=0.5, shape=(2,2)) = [[ 7.10486984,  3.37695289],
                                         [ 3.91697288,  3.65933681]]

Defined in src/operator/random/sample_op.cc:L125

Parameters:
  • alpha (float, optional, default=1) – Alpha parameter (shape) of the gamma distribution.
  • beta (float, optional, default=1) – Beta parameter (scale) of the gamma distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_generalized_negative_binomial(mu=_Null, alpha=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a generalized negative binomial distribution.

Samples are distributed according to a generalized negative binomial distribution parametrized by mu (mean) and alpha (dispersion). alpha is defined as 1/k where k is the failure limit of the number of unsuccessful experiments (generalized to real numbers). Samples will always be returned as a floating point data type.

Example:

generalized_negative_binomial(mu=2.0, alpha=0.3, shape=(2,2)) = [[ 2.,  1.],
                                                                 [ 6.,  4.]]

Defined in src/operator/random/sample_op.cc:L179

Parameters:
  • mu (float, optional, default=1) – Mean of the negative binomial distribution.
  • alpha (float, optional, default=1) – Alpha (dispersion) parameter of the negative binomial distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_negative_binomial(k=_Null, p=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a negative binomial distribution.

Samples are distributed according to a negative binomial distribution parametrized by k (limit of unsuccessful experiments) and p (failure probability in each experiment). Samples will always be returned as a floating point data type.

Example:

negative_binomial(k=3, p=0.4, shape=(2,2)) = [[ 4.,  7.],
                                              [ 2.,  5.]]

Defined in src/operator/random/sample_op.cc:L164

Parameters:
  • k (int, optional, default='1') – Limit of unsuccessful experiments.
  • p (float, optional, default=1) – Failure probability in each experiment.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_normal(loc=_Null, scale=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a normal (Gaussian) distribution.

Note

The existing alias normal is deprecated.

Samples are distributed according to a normal distribution parametrized by loc (mean) and scale (standard deviation).

Example:

normal(loc=0, scale=1, shape=(2,2)) = [[ 1.89171135, -1.16881478],
                                       [-1.23474145,  1.55807114]]

Defined in src/operator/random/sample_op.cc:L113

Parameters:
  • loc (float, optional, default=0) – Mean of the distribution.
  • scale (float, optional, default=1) – Standard deviation of the distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_poisson(lam=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a Poisson distribution.

Samples are distributed according to a Poisson distribution parametrized by lambda (rate). Samples will always be returned as a floating point data type.

Example:

poisson(lam=4, shape=(2,2)) = [[ 5.,  2.],
                               [ 4.,  6.]]

Defined in src/operator/random/sample_op.cc:L150

Parameters:
  • lam (float, optional, default=1) – Lambda parameter (rate) of the Poisson distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_randint(low=_Null, high=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a discrete uniform distribution.

Samples are uniformly distributed over the half-open interval [low, high) (includes low, but excludes high).

Example:

randint(low=0, high=5, shape=(2,2)) = [[ 0,  2],
                                       [ 3,  1]]

Defined in src/operator/random/sample_op.cc:L193

Parameters:
  • low (, required) – Lower bound of the distribution.
  • high (, required) – Upper bound of the distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'int32', 'int64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to int32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.random_uniform(low=_Null, high=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a uniform distribution.

Note

The existing alias uniform is deprecated.

Samples are uniformly distributed over the half-open interval [low, high) (includes low, but excludes high).

Example:

uniform(low=0, high=1, shape=(2,2)) = [[ 0.60276335,  0.85794562],
                                       [ 0.54488319,  0.84725171]]

Defined in src/operator/random/sample_op.cc:L96

Parameters:
  • low (float, optional, default=0) – Lower bound of the distribution.
  • high (float, optional, default=1) – Upper bound of the distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.ravel_multi_index(data=None, shape=_Null, out=None, name=None, **kwargs)

Converts a batch of index arrays into an array of flat indices. The operator follows numpy conventions so a single multi index is given by a column of the input matrix. The leading dimension may be left unspecified by using -1 as placeholder.

Examples:

A = [[3,6,6],[4,5,1]]
ravel(A, shape=(7,6)) = [22,41,37]
ravel(A, shape=(-1,6)) = [22,41,37]

Defined in src/operator/tensor/ravel.cc:L42

Parameters:
  • data (NDArray) – Batch of multi-indices
  • shape (Shape(tuple), optional, default=None) – Shape of the array into which the multi-indices apply.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.rcbrt(data=None, out=None, name=None, **kwargs)

Returns element-wise inverse cube-root value of the input.

\[rcbrt(x) = 1/\sqrt[3]{x}\]

Example:

rcbrt([1,8,-125]) = [1.0, 0.5, -0.2]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L1004

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.reciprocal(data=None, out=None, name=None, **kwargs)

Returns the reciprocal of the argument, element-wise.

Calculates 1/x.

Example:

reciprocal([-2, 1, 3, 1.6.0, 0.2]) = [-0.5, 1.0, 0.33333334, 0.625, 5.0]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L686

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.relu(data=None, out=None, name=None, **kwargs)

Computes rectified linear activation.

\[max(features, 0)\]

The storage type of relu output depends upon the input storage type:

  • relu(default) = default
  • relu(row_sparse) = row_sparse
  • relu(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L85

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.repeat(data=None, repeats=_Null, axis=_Null, out=None, name=None, **kwargs)

Repeats elements of an array.

By default, repeat flattens the input array into 1-D and then repeats the elements:

x = [[ 1, 2],
     [ 3, 4]]

repeat(x, repeats=2) = [ 1.,  1.,  2.,  2.,  3.,  3.,  4.,  4.]

The parameter axis specifies the axis along which to perform repeat:

repeat(x, repeats=2, axis=1) = [[ 1.,  1.,  2.,  2.],
                                [ 3.,  3.,  4.,  4.]]

repeat(x, repeats=2, axis=0) = [[ 1.,  2.],
                                [ 1.,  2.],
                                [ 3.,  4.],
                                [ 3.,  4.]]

repeat(x, repeats=2, axis=-1) = [[ 1.,  1.,  2.,  2.],
                                 [ 3.,  3.,  4.,  4.]]

Defined in src/operator/tensor/matrix_op.cc:L796

Parameters:
  • data (NDArray) – Input data array
  • repeats (int, required) – The number of repetitions for each element.
  • axis (int or None, optional, default='None') – The axis along which to repeat values. The negative numbers are interpreted counting from the backward. By default, use the flattened input array, and return a flat output array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.reshape(data=None, shape=_Null, reverse=_Null, target_shape=_Null, keep_highest=_Null, out=None, name=None, **kwargs)

Reshapes the input array.

Note

Reshape is deprecated, use reshape

Given an array and a shape, this function returns a copy of the array in the new shape. The shape is a tuple of integers such as (2,3,4). The size of the new shape should be same as the size of the input array.

Example:

reshape([1,2,3,4], shape=(2,2)) = [[1,2], [3,4]]

Some dimensions of the shape can take special values from the set {0, -1, -2, -3, -4}. The significance of each is explained below:

  • 0 copy this dimension from the input to the output shape.

    Example:

    - input shape = (2,3,4), shape = (4,0,2), output shape = (4,3,2)
    - input shape = (2,3,4), shape = (2,0,0), output shape = (2,3,4)
    
  • -1 infers the dimension of the output shape by using the remainder of the input dimensions keeping the size of the new array same as that of the input array. At most one dimension of shape can be -1.

    Example:

    - input shape = (2,3,4), shape = (6,1,-1), output shape = (6,1,4)
    - input shape = (2,3,4), shape = (3,-1,8), output shape = (3,1,8)
    - input shape = (2,3,4), shape=(-1,), output shape = (24,)
    
  • -2 copy all/remainder of the input dimensions to the output shape.

    Example:

    - input shape = (2,3,4), shape = (-2,), output shape = (2,3,4)
    - input shape = (2,3,4), shape = (2,-2), output shape = (2,3,4)
    - input shape = (2,3,4), shape = (-2,1,1), output shape = (2,3,4,1,1)
    
  • -3 use the product of two consecutive dimensions of the input shape as the output dimension.

    Example:

    - input shape = (2,3,4), shape = (-3,4), output shape = (6,4)
    - input shape = (2,3,4,5), shape = (-3,-3), output shape = (6,20)
    - input shape = (2,3,4), shape = (0,-3), output shape = (2,12)
    - input shape = (2,3,4), shape = (-3,-2), output shape = (6,4)
    
  • -4 split one dimension of the input into two dimensions passed subsequent to -4 in shape (can contain -1).

    Example:

    - input shape = (2,3,4), shape = (-4,1,2,-2), output shape =(1,2,3,4)
    - input shape = (2,3,4), shape = (2,-4,-1,3,-2), output shape = (2,1,3,4)
    

If the argument reverse is set to 1, then the special values are inferred from right to left.

Example:

- without reverse=1, for input shape = (10,5,4), shape = (-1,0), output shape would be (40,5)
- with reverse=1, output shape will be (50,4).

Defined in src/operator/tensor/matrix_op.cc:L202

Parameters:
  • data (NDArray) – Input data to reshape.
  • shape (Shape(tuple), optional, default=[]) – The target shape
  • reverse (boolean, optional, default=0) – If true then the special values are inferred from right to left
  • target_shape (Shape(tuple), optional, default=[]) – (Deprecated! Use shape instead.) Target new shape. One and only one dim can be 0, in which case it will be inferred from the rest of dims
  • keep_highest (boolean, optional, default=0) – (Deprecated! Use shape instead.) Whether keep the highest dim unchanged.If set to true, then the first dim in target_shape is ignored,and always fixed as input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.reshape_like(lhs=None, rhs=None, out=None, name=None, **kwargs)

Reshape some or all dimensions of lhs to have the same shape as some or all dimensions of rhs.

Returns a view of the lhs array with a new shape without altering any data.

Example:

x = [1, 2, 3, 4, 5, 6]
y = [[0, -4], [3, 2], [2, 2]]
reshape_like(x, y) = [[1, 2], [3, 4], [5, 6]]

More precise control over how dimensions are inherited is achieved by specifying slices over the lhs and rhs array dimensions. Only the sliced lhs dimensions are reshaped to the rhs sliced dimensions, with the non-sliced lhs dimensions staying the same.

Examples:

- lhs shape = (30,7), rhs shape = (15,2,4), lhs_begin=0, lhs_end=1, rhs_begin=0, rhs_end=2, output shape = (15,2,7)
- lhs shape = (3, 5), rhs shape = (1,15,4), lhs_begin=0, lhs_end=2, rhs_begin=1, rhs_end=2, output shape = (15)

Negative indices are supported, and None can be used for either lhs_end or rhs_end to indicate the end of the range.

Example:

- lhs shape = (30, 12), rhs shape = (4, 2, 2, 3), lhs_begin=-1, lhs_end=None, rhs_begin=1, rhs_end=None, output shape = (30, 2, 2, 3)

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L485

Parameters:
  • lhs (NDArray) – First input.
  • rhs (NDArray) – Second input.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.reverse(data=None, axis=_Null, out=None, name=None, **kwargs)

Reverses the order of elements along given axis while preserving array shape.

Note: reverse and flip are equivalent. We use reverse in the following examples.

Examples:

x = [[ 0.,  1.,  2.,  3.,  4.],
     [ 5.,  6.,  7.,  8.,  9.]]

reverse(x, axis=0) = [[ 5.,  6.,  7.,  8.,  9.],
                      [ 0.,  1.,  2.,  3.,  4.]]

reverse(x, axis=1) = [[ 4.,  3.,  2.,  1.,  0.],
                      [ 9.,  8.,  7.,  6.,  5.]]

Defined in src/operator/tensor/matrix_op.cc:L898

Parameters:
  • data (NDArray) – Input data array
  • axis (Shape(tuple), required) – The axis which to reverse elements.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.rint(data=None, out=None, name=None, **kwargs)

Returns element-wise rounded value to the nearest integer of the input.

Note

  • For input n.5 rint returns n while round returns n+1.
  • For input -n.5 both rint and round returns -n-1.

Example:

rint([-1.5, 1.5, -1.9, 1.9, 2.1]) = [-2.,  1., -2.,  2.,  2.]

The storage type of rint output depends upon the input storage type:

  • rint(default) = default
  • rint(row_sparse) = row_sparse
  • rint(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L767

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.rmsprop_update(weight=None, grad=None, n=None, lr=_Null, gamma1=_Null, epsilon=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, clip_weights=_Null, out=None, name=None, **kwargs)

Update function for RMSProp optimizer.

RMSprop is a variant of stochastic gradient descent where the gradients are divided by a cache which grows with the sum of squares of recent gradients?

RMSProp is similar to AdaGrad, a popular variant of SGD which adaptively tunes the learning rate of each parameter. AdaGrad lowers the learning rate for each parameter monotonically over the course of training. While this is analytically motivated for convex optimizations, it may not be ideal for non-convex problems. RMSProp deals with this heuristically by allowing the learning rates to rebound as the denominator decays over time.

Define the Root Mean Square (RMS) error criterion of the gradient as \(RMS[g]_t = \sqrt{E[g^2]_t + \epsilon}\), where \(g\) represents gradient and \(E[g^2]_t\) is the decaying average over past squared gradient.

The \(E[g^2]_t\) is given by:

\[E[g^2]_t = \gamma * E[g^2]_{t-1} + (1-\gamma) * g_t^2\]

The update step is

\[\theta_{t+1} = \theta_t - \frac{\eta}{RMS[g]_t} g_t\]

The RMSProp code follows the version in http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf Tieleman & Hinton, 2012.

Hinton suggests the momentum term \(\gamma\) to be 0.9 and the learning rate \(\eta\) to be 0.001.

Defined in src/operator/optimizer_op.cc:L795

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • n (NDArray) – n
  • lr (float, required) – Learning rate
  • gamma1 (float, optional, default=0.949999988) – The decay rate of momentum estimates.
  • epsilon (float, optional, default=9.99999994e-09) – A small constant for numerical stability.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • clip_weights (float, optional, default=-1) – Clip weights to the range of [-clip_weights, clip_weights] If clip_weights <= 0, weight clipping is turned off. weights = max(min(weights, clip_weights), -clip_weights).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.rmspropalex_update(weight=None, grad=None, n=None, g=None, delta=None, lr=_Null, gamma1=_Null, gamma2=_Null, epsilon=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, clip_weights=_Null, out=None, name=None, **kwargs)

Update function for RMSPropAlex optimizer.

RMSPropAlex is non-centered version of RMSProp.

Define \(E[g^2]_t\) is the decaying average over past squared gradient and \(E[g]_t\) is the decaying average over past gradient.

\[\begin{split}E[g^2]_t = \gamma_1 * E[g^2]_{t-1} + (1 - \gamma_1) * g_t^2\\ E[g]_t = \gamma_1 * E[g]_{t-1} + (1 - \gamma_1) * g_t\\ \Delta_t = \gamma_2 * \Delta_{t-1} - \frac{\eta}{\sqrt{E[g^2]_t - E[g]_t^2 + \epsilon}} g_t\\\end{split}\]

The update step is

\[\theta_{t+1} = \theta_t + \Delta_t\]

The RMSPropAlex code follows the version in http://arxiv.org/pdf/1308.0850v5.pdf Eq(38) - Eq(45) by Alex Graves, 2013.

Graves suggests the momentum term \(\gamma_1\) to be 0.95, \(\gamma_2\) to be 0.9 and the learning rate \(\eta\) to be 0.0001.

Defined in src/operator/optimizer_op.cc:L834

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • n (NDArray) – n
  • g (NDArray) – g
  • delta (NDArray) – delta
  • lr (float, required) – Learning rate
  • gamma1 (float, optional, default=0.949999988) – Decay rate.
  • gamma2 (float, optional, default=0.899999976) – Decay rate.
  • epsilon (float, optional, default=9.99999994e-09) – A small constant for numerical stability.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • clip_weights (float, optional, default=-1) – Clip weights to the range of [-clip_weights, clip_weights] If clip_weights <= 0, weight clipping is turned off. weights = max(min(weights, clip_weights), -clip_weights).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.round(data=None, out=None, name=None, **kwargs)

Returns element-wise rounded value to the nearest integer of the input.

Example:

round([-1.5, 1.5, -1.9, 1.9, 2.1]) = [-2.,  2., -2.,  2.,  2.]

The storage type of round output depends upon the input storage type:

  • round(default) = default
  • round(row_sparse) = row_sparse
  • round(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L746

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.rsqrt(data=None, out=None, name=None, **kwargs)

Returns element-wise inverse square-root value of the input.

\[rsqrt(x) = 1/\sqrt{x}\]

Example:

rsqrt([4,9,16]) = [0.5, 0.33333334, 0.25]

The storage type of rsqrt output is always dense

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L927

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_exponential(lam=None, shape=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple exponential distributions with parameters lambda (rate).

The parameters of the distributions are provided as an input array. Let [s] be the shape of the input array, n be the dimension of [s], [t] be the shape specified as the parameter of the operator, and m be the dimension of [t]. Then the output will be a (n+m)-dimensional array with shape [s]x[t].

For any valid n-dimensional index i with respect to the input array, output[i] will be an m-dimensional array that holds randomly drawn samples from the distribution which is parameterized by the input value at index i. If the shape parameter of the operator is not set, then one sample will be drawn per distribution and the output array has the same shape as the input array.

Examples:

lam = [ 1.0, 8.5 ]

// Draw a single sample for each distribution
sample_exponential(lam) = [ 0.51837951,  0.09994757]

// Draw a vector containing two samples for each distribution
sample_exponential(lam, shape=(2)) = [[ 0.51837951,  0.19866663],
                                      [ 0.09994757,  0.50447971]]

Defined in src/operator/random/multisample_op.cc:L284

Parameters:
  • lam (NDArray) – Lambda (rate) parameters of the distributions.
  • shape (Shape(tuple), optional, default=None) – Shape to be sampled from each random distribution.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_gamma(alpha=None, beta=None, shape=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple gamma distributions with parameters alpha (shape) and beta (scale).

The parameters of the distributions are provided as input arrays. Let [s] be the shape of the input arrays, n be the dimension of [s], [t] be the shape specified as the parameter of the operator, and m be the dimension of [t]. Then the output will be a (n+m)-dimensional array with shape [s]x[t].

For any valid n-dimensional index i with respect to the input arrays, output[i] will be an m-dimensional array that holds randomly drawn samples from the distribution which is parameterized by the input values at index i. If the shape parameter of the operator is not set, then one sample will be drawn per distribution and the output array has the same shape as the input arrays.

Examples:

alpha = [ 0.0, 2.5 ]
beta = [ 1.0, 0.7 ]

// Draw a single sample for each distribution
sample_gamma(alpha, beta) = [ 0.        ,  2.25797319]

// Draw a vector containing two samples for each distribution
sample_gamma(alpha, beta, shape=(2)) = [[ 0.        ,  0.        ],
                                        [ 2.25797319,  1.70734084]]

Defined in src/operator/random/multisample_op.cc:L282

Parameters:
  • alpha (NDArray) – Alpha (shape) parameters of the distributions.
  • shape (Shape(tuple), optional, default=None) – Shape to be sampled from each random distribution.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • beta (NDArray) – Beta (scale) parameters of the distributions.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_generalized_negative_binomial(mu=None, alpha=None, shape=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple generalized negative binomial distributions with parameters mu (mean) and alpha (dispersion).

The parameters of the distributions are provided as input arrays. Let [s] be the shape of the input arrays, n be the dimension of [s], [t] be the shape specified as the parameter of the operator, and m be the dimension of [t]. Then the output will be a (n+m)-dimensional array with shape [s]x[t].

For any valid n-dimensional index i with respect to the input arrays, output[i] will be an m-dimensional array that holds randomly drawn samples from the distribution which is parameterized by the input values at index i. If the shape parameter of the operator is not set, then one sample will be drawn per distribution and the output array has the same shape as the input arrays.

Samples will always be returned as a floating point data type.

Examples:

mu = [ 2.0, 2.5 ]
alpha = [ 1.0, 0.1 ]

// Draw a single sample for each distribution
sample_generalized_negative_binomial(mu, alpha) = [ 0.,  3.]

// Draw a vector containing two samples for each distribution
sample_generalized_negative_binomial(mu, alpha, shape=(2)) = [[ 0.,  3.],
                                                              [ 3.,  1.]]

Defined in src/operator/random/multisample_op.cc:L293

Parameters:
  • mu (NDArray) – Means of the distributions.
  • shape (Shape(tuple), optional, default=None) – Shape to be sampled from each random distribution.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • alpha (NDArray) – Alpha (dispersion) parameters of the distributions.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_multinomial(data=None, shape=_Null, get_prob=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple multinomial distributions.

data is an n dimensional array whose last dimension has length k, where k is the number of possible outcomes of each multinomial distribution. This operator will draw shape samples from each distribution. If shape is empty one sample will be drawn from each distribution.

If get_prob is true, a second array containing log likelihood of the drawn samples will also be returned. This is usually used for reinforcement learning where you can provide reward as head gradient for this array to estimate gradient.

Note that the input distribution must be normalized, i.e. data must sum to 1 along its last axis.

Examples:

probs = [[0, 0.1, 0.2, 0.3, 0.4], [0.4, 0.3, 0.2, 0.1, 0]]

// Draw a single sample for each distribution
sample_multinomial(probs) = [3, 0]

// Draw a vector containing two samples for each distribution
sample_multinomial(probs, shape=(2)) = [[4, 2],
                                        [0, 0]]

// requests log likelihood
sample_multinomial(probs, get_prob=True) = [2, 1], [0.2, 0.3]
Parameters:
  • data (NDArray) – Distribution probabilities. Must sum to one on the last axis.
  • shape (Shape(tuple), optional, default=[]) – Shape to be sampled from each random distribution.
  • get_prob (boolean, optional, default=0) – Whether to also return the log probability of sampled result. This is usually used for differentiating through stochastic variables, e.g. in reinforcement learning.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'uint8'},optional, default='int32') – DType of the output in case this can’t be inferred.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_negative_binomial(k=None, p=None, shape=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple negative binomial distributions with parameters k (failure limit) and p (failure probability).

The parameters of the distributions are provided as input arrays. Let [s] be the shape of the input arrays, n be the dimension of [s], [t] be the shape specified as the parameter of the operator, and m be the dimension of [t]. Then the output will be a (n+m)-dimensional array with shape [s]x[t].

For any valid n-dimensional index i with respect to the input arrays, output[i] will be an m-dimensional array that holds randomly drawn samples from the distribution which is parameterized by the input values at index i. If the shape parameter of the operator is not set, then one sample will be drawn per distribution and the output array has the same shape as the input arrays.

Samples will always be returned as a floating point data type.

Examples:

k = [ 20, 49 ]
p = [ 0.4 , 0.77 ]

// Draw a single sample for each distribution
sample_negative_binomial(k, p) = [ 15.,  16.]

// Draw a vector containing two samples for each distribution
sample_negative_binomial(k, p, shape=(2)) = [[ 15.,  50.],
                                             [ 16.,  12.]]

Defined in src/operator/random/multisample_op.cc:L289

Parameters:
  • k (NDArray) – Limits of unsuccessful experiments.
  • shape (Shape(tuple), optional, default=None) – Shape to be sampled from each random distribution.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • p (NDArray) – Failure probabilities in each experiment.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_normal(mu=None, sigma=None, shape=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple normal distributions with parameters mu (mean) and sigma (standard deviation).

The parameters of the distributions are provided as input arrays. Let [s] be the shape of the input arrays, n be the dimension of [s], [t] be the shape specified as the parameter of the operator, and m be the dimension of [t]. Then the output will be a (n+m)-dimensional array with shape [s]x[t].

For any valid n-dimensional index i with respect to the input arrays, output[i] will be an m-dimensional array that holds randomly drawn samples from the distribution which is parameterized by the input values at index i. If the shape parameter of the operator is not set, then one sample will be drawn per distribution and the output array has the same shape as the input arrays.

Examples:

mu = [ 0.0, 2.5 ]
sigma = [ 1.0, 3.7 ]

// Draw a single sample for each distribution
sample_normal(mu, sigma) = [-0.56410581,  0.95934606]

// Draw a vector containing two samples for each distribution
sample_normal(mu, sigma, shape=(2)) = [[-0.56410581,  0.2928229 ],
                                       [ 0.95934606,  4.48287058]]

Defined in src/operator/random/multisample_op.cc:L279

Parameters:
  • mu (NDArray) – Means of the distributions.
  • shape (Shape(tuple), optional, default=None) – Shape to be sampled from each random distribution.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • sigma (NDArray) – Standard deviations of the distributions.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_poisson(lam=None, shape=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple Poisson distributions with parameters lambda (rate).

The parameters of the distributions are provided as an input array. Let [s] be the shape of the input array, n be the dimension of [s], [t] be the shape specified as the parameter of the operator, and m be the dimension of [t]. Then the output will be a (n+m)-dimensional array with shape [s]x[t].

For any valid n-dimensional index i with respect to the input array, output[i] will be an m-dimensional array that holds randomly drawn samples from the distribution which is parameterized by the input value at index i. If the shape parameter of the operator is not set, then one sample will be drawn per distribution and the output array has the same shape as the input array.

Samples will always be returned as a floating point data type.

Examples:

lam = [ 1.0, 8.5 ]

// Draw a single sample for each distribution
sample_poisson(lam) = [  0.,  13.]

// Draw a vector containing two samples for each distribution
sample_poisson(lam, shape=(2)) = [[  0.,   4.],
                                  [ 13.,   8.]]

Defined in src/operator/random/multisample_op.cc:L286

Parameters:
  • lam (NDArray) – Lambda (rate) parameters of the distributions.
  • shape (Shape(tuple), optional, default=None) – Shape to be sampled from each random distribution.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sample_uniform(low=None, high=None, shape=_Null, dtype=_Null, out=None, name=None, **kwargs)

Concurrent sampling from multiple uniform distributions on the intervals given by [low,high).

The parameters of the distributions are provided as input arrays. Let [s] be the shape of the input arrays, n be the dimension of [s], [t] be the shape specified as the parameter of the operator, and m be the dimension of [t]. Then the output will be a (n+m)-dimensional array with shape [s]x[t].

For any valid n-dimensional index i with respect to the input arrays, output[i] will be an m-dimensional array that holds randomly drawn samples from the distribution which is parameterized by the input values at index i. If the shape parameter of the operator is not set, then one sample will be drawn per distribution and the output array has the same shape as the input arrays.

Examples:

low = [ 0.0, 2.5 ]
high = [ 1.0, 3.7 ]

// Draw a single sample for each distribution
sample_uniform(low, high) = [ 0.40451524,  3.18687344]

// Draw a vector containing two samples for each distribution
sample_uniform(low, high, shape=(2)) = [[ 0.40451524,  0.18017688],
                                        [ 3.18687344,  3.68352246]]

Defined in src/operator/random/multisample_op.cc:L277

Parameters:
  • low (NDArray) – Lower bounds of the distributions.
  • shape (Shape(tuple), optional, default=None) – Shape to be sampled from each random distribution.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • high (NDArray) – Upper bounds of the distributions.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.scatter_nd(data=None, indices=None, shape=_Null, out=None, name=None, **kwargs)

Scatters data into a new tensor according to indices.

Given data with shape (Y_0, ..., Y_{K-1}, X_M, ..., X_{N-1}) and indices with shape (M, Y_0, ..., Y_{K-1}), the output will have shape (X_0, X_1, ..., X_{N-1}), where M <= N. If M == N, data shape should simply be (Y_0, ..., Y_{K-1}).

The elements in output is defined as follows:

output[indices[0, y_0, ..., y_{K-1}],
       ...,
       indices[M-1, y_0, ..., y_{K-1}],
       x_M, ..., x_{N-1}] = data[y_0, ..., y_{K-1}, x_M, ..., x_{N-1}]

all other entries in output are 0.

Warning

If the indices have duplicates, the result will be non-deterministic and the gradient of scatter_nd will not be correct!!

Examples:

data = [2, 3, 0]
indices = [[1, 1, 0], [0, 1, 0]]
shape = (2, 2)
scatter_nd(data, indices, shape) = [[0, 0], [2, 3]]

data = [[[1, 2], [3, 4]], [[5, 6], [7, 8]]]
indices = [[0, 1], [1, 1]]
shape = (2, 2, 2, 2)
scatter_nd(data, indices, shape) = [[[[0, 0],
                                      [0, 0]],

                                     [[1, 2],
                                      [3, 4]]],

                                    [[[0, 0],
                                      [0, 0]],

                                     [[5, 6],
                                      [7, 8]]]]
Parameters:
  • data (NDArray) – data
  • indices (NDArray) – indices
  • shape (Shape(tuple), required) – Shape of output.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sgd_mom_update(weight=None, grad=None, mom=None, lr=_Null, momentum=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, lazy_update=_Null, out=None, name=None, **kwargs)

Momentum update function for Stochastic Gradient Descent (SGD) optimizer.

Momentum update has better convergence rates on neural networks. Mathematically it looks like below:

\[\begin{split}v_1 = \alpha * \nabla J(W_0)\\ v_t = \gamma v_{t-1} - \alpha * \nabla J(W_{t-1})\\ W_t = W_{t-1} + v_t\end{split}\]

It updates the weights using:

v = momentum * v - learning_rate * gradient
weight += v

Where the parameter momentum is the decay rate of momentum estimates at each epoch.

However, if grad’s storage type is row_sparse, lazy_update is True and weight’s storage type is the same as momentum’s storage type, only the row slices whose indices appear in grad.indices are updated (for both weight and momentum):

for row in gradient.indices:
    v[row] = momentum[row] * v[row] - learning_rate * gradient[row]
    weight[row] += v[row]

Defined in src/operator/optimizer_op.cc:L563

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • mom (NDArray) – Momentum
  • lr (float, required) – Learning rate
  • momentum (float, optional, default=0) – The decay rate of momentum estimates at each epoch.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • lazy_update (boolean, optional, default=1) – If true, lazy updates are applied if gradient’s stype is row_sparse and both weight and momentum have the same stype
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sgd_update(weight=None, grad=None, lr=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, lazy_update=_Null, out=None, name=None, **kwargs)

Update function for Stochastic Gradient Descent (SGD) optimizer.

It updates the weights using:

weight = weight - learning_rate * (gradient + wd * weight)

However, if gradient is of row_sparse storage type and lazy_update is True, only the row slices whose indices appear in grad.indices are updated:

for row in gradient.indices:
    weight[row] = weight[row] - learning_rate * (gradient[row] + wd * weight[row])

Defined in src/operator/optimizer_op.cc:L522

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • lr (float, required) – Learning rate
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • lazy_update (boolean, optional, default=1) – If true, lazy updates are applied if gradient’s stype is row_sparse.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.shape_array(data=None, lhs_begin=_Null, lhs_end=_Null, rhs_begin=_Null, rhs_end=_Null, out=None, name=None, **kwargs)

Returns a 1D int64 array containing the shape of data.

Example:

shape_array([[1,2,3,4], [5,6,7,8]]) = [2,4]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L544

Parameters:
  • data (NDArray) – Input Array.
  • lhs_begin (int or None, optional, default='None') – Defaults to 0. The beginning index along which the lhs dimensions are to be reshaped. Supports negative indices.
  • lhs_end (int or None, optional, default='None') – Defaults to None. The ending index along which the lhs dimensions are to be used for reshaping. Supports negative indices.
  • rhs_begin (int or None, optional, default='None') – Defaults to 0. The beginning index along which the rhs dimensions are to be used for reshaping. Supports negative indices.
  • rhs_end (int or None, optional, default='None') – Defaults to None. The ending index along which the rhs dimensions are to be used for reshaping. Supports negative indices.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.shuffle(data=None, out=None, name=None, **kwargs)

Randomly shuffle the elements.

This shuffles the array along the first axis. The order of the elements in each subarray does not change. For example, if a 2D array is given, the order of the rows randomly changes, but the order of the elements in each row does not change.

Parameters:
  • data (NDArray) – Data to be shuffled.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sigmoid(data=None, out=None, name=None, **kwargs)

Computes sigmoid of x element-wise.

\[y = 1 / (1 + exp(-x))\]

The storage type of sigmoid output is always dense

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L119

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sign(data=None, out=None, name=None, **kwargs)

Returns element-wise sign of the input.

Example:

sign([-2, 0, 3]) = [-1, 0, 1]

The storage type of sign output depends upon the input storage type:

  • sign(default) = default
  • sign(row_sparse) = row_sparse
  • sign(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L727

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.signsgd_update(weight=None, grad=None, lr=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, out=None, name=None, **kwargs)

Update function for SignSGD optimizer.

\[\begin{split}g_t = \nabla J(W_{t-1})\\ W_t = W_{t-1} - \eta_t \text{sign}(g_t)\end{split}\]

It updates the weights using:

weight = weight - learning_rate * sign(gradient)

Note

  • sparse ndarray not supported for this optimizer yet.

Defined in src/operator/optimizer_op.cc:L61

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • lr (float, required) – Learning rate
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.signum_update(weight=None, grad=None, mom=None, lr=_Null, momentum=_Null, wd=_Null, rescale_grad=_Null, clip_gradient=_Null, wd_lh=_Null, out=None, name=None, **kwargs)

SIGN momentUM (Signum) optimizer.

\[\begin{split}g_t = \nabla J(W_{t-1})\\ m_t = \beta m_{t-1} + (1 - \beta) g_t\\ W_t = W_{t-1} - \eta_t \text{sign}(m_t)\end{split}\]
It updates the weights using::
state = momentum * state + (1-momentum) * gradient weight = weight - learning_rate * sign(state)

Where the parameter momentum is the decay rate of momentum estimates at each epoch.

Note

  • sparse ndarray not supported for this optimizer yet.

Defined in src/operator/optimizer_op.cc:L90

Parameters:
  • weight (NDArray) – Weight
  • grad (NDArray) – Gradient
  • mom (NDArray) – Momentum
  • lr (float, required) – Learning rate
  • momentum (float, optional, default=0) – The decay rate of momentum estimates at each epoch.
  • wd (float, optional, default=0) – Weight decay augments the objective function with a regularization term that penalizes large weights. The penalty scales with the square of the magnitude of each weight.
  • rescale_grad (float, optional, default=1) – Rescale gradient to grad = rescale_grad*grad.
  • clip_gradient (float, optional, default=-1) – Clip gradient to the range of [-clip_gradient, clip_gradient] If clip_gradient <= 0, gradient clipping is turned off. grad = max(min(grad, clip_gradient), -clip_gradient).
  • wd_lh (float, optional, default=0) – The amount of weight decay that does not go into gradient/momentum calculationsotherwise do weight decay algorithmically only.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sin(data=None, out=None, name=None, **kwargs)

Computes the element-wise sine of the input array.

The input should be in radians (\(2\pi\) rad equals 360 degrees).

\[sin([0, \pi/4, \pi/2]) = [0, 0.707, 1]\]

The storage type of sin output depends upon the input storage type:

  • sin(default) = default
  • sin(row_sparse) = row_sparse
  • sin(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L46

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sinh(data=None, out=None, name=None, **kwargs)

Returns the hyperbolic sine of the input array, computed element-wise.

\[sinh(x) = 0.5\times(exp(x) - exp(-x))\]

The storage type of sinh output depends upon the input storage type:

  • sinh(default) = default
  • sinh(row_sparse) = row_sparse
  • sinh(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L257

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.size_array(data=None, out=None, name=None, **kwargs)

Returns a 1D int64 array containing the size of data.

Example:

size_array([[1,2,3,4], [5,6,7,8]]) = [8]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L596

Parameters:
  • data (NDArray) – Input Array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.slice(data=None, begin=_Null, end=_Null, step=_Null, out=None, name=None, **kwargs)

Slices a region of the array.

Note

crop is deprecated. Use slice instead.

This function returns a sliced array between the indices given by begin and end with the corresponding step.

For an input array of shape=(d_0, d_1, ..., d_n-1), slice operation with begin=(b_0, b_1...b_m-1), end=(e_0, e_1, ..., e_m-1), and step=(s_0, s_1, ..., s_m-1), where m <= n, results in an array with the shape (|e_0-b_0|/|s_0|, ..., |e_m-1-b_m-1|/|s_m-1|, d_m, ..., d_n-1).

The resulting array’s k-th dimension contains elements from the k-th dimension of the input array starting from index b_k (inclusive) with step s_k until reaching e_k (exclusive).

If the k-th elements are None in the sequence of begin, end, and step, the following rule will be used to set default values. If s_k is None, set s_k=1. If s_k > 0, set b_k=0, e_k=d_k; else, set b_k=d_k-1, e_k=-1.

The storage type of slice output depends on storage types of inputs

  • slice(csr) = csr
  • otherwise, slice generates output with default storage

Note

When input data storage type is csr, it only supports step=(), or step=(None,), or step=(1,) to generate a csr output. For other step parameter values, it falls back to slicing a dense tensor.

Example:

x = [[  1.,   2.,   3.,   4.],
     [  5.,   6.,   7.,   8.],
     [  9.,  10.,  11.,  12.]]

slice(x, begin=(0,1), end=(2,4)) = [[ 2.,  3.,  4.],
                                   [ 6.,  7.,  8.]]
slice(x, begin=(None, 0), end=(None, 3), step=(-1, 2)) = [[9., 11.],
                                                          [5.,  7.],
                                                          [1.,  3.]]

Defined in src/operator/tensor/matrix_op.cc:L506

Parameters:
  • data (NDArray) – Source input
  • begin (Shape(tuple), required) – starting indices for the slice operation, supports negative indices.
  • end (Shape(tuple), required) – ending indices for the slice operation, supports negative indices.
  • step (Shape(tuple), optional, default=[]) – step for the slice operation, supports negative values.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.slice_axis(data=None, axis=_Null, begin=_Null, end=_Null, out=None, name=None, **kwargs)

Slices along a given axis.

Returns an array slice along a given axis starting from the begin index to the end index.

Examples:

x = [[  1.,   2.,   3.,   4.],
     [  5.,   6.,   7.,   8.],
     [  9.,  10.,  11.,  12.]]

slice_axis(x, axis=0, begin=1, end=3) = [[  5.,   6.,   7.,   8.],
                                         [  9.,  10.,  11.,  12.]]

slice_axis(x, axis=1, begin=0, end=2) = [[  1.,   2.],
                                         [  5.,   6.],
                                         [  9.,  10.]]

slice_axis(x, axis=1, begin=-3, end=-1) = [[  2.,   3.],
                                           [  6.,   7.],
                                           [ 10.,  11.]]

Defined in src/operator/tensor/matrix_op.cc:L596

Parameters:
  • data (NDArray) – Source input
  • axis (int, required) – Axis along which to be sliced, supports negative indexes.
  • begin (int, required) – The beginning index along the axis to be sliced, supports negative indexes.
  • end (int or None, required) – The ending index along the axis to be sliced, supports negative indexes.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.slice_like(data=None, shape_like=None, axes=_Null, out=None, name=None, **kwargs)

Slices a region of the array like the shape of another array.

This function is similar to slice, however, the begin are always 0`s and `end of specific axes are inferred from the second input shape_like.

Given the second shape_like input of shape=(d_0, d_1, ..., d_n-1), a slice_like operator with default empty axes, it performs the following operation:

`` out = slice(input, begin=(0, 0, ..., 0), end=(d_0, d_1, ..., d_n-1))``.

When axes is not empty, it is used to speficy which axes are being sliced.

Given a 4-d input data, slice_like operator with axes=(0, 2, -1) will perform the following operation:

`` out = slice(input, begin=(0, 0, 0, 0), end=(d_0, None, d_2, d_3))``.

Note that it is allowed to have first and second input with different dimensions, however, you have to make sure the axes are specified and not exceeding the dimension limits.

For example, given input_1 with shape=(2,3,4,5) and input_2 with shape=(1,2,3), it is not allowed to use:

`` out = slice_like(a, b)`` because ndim of input_1 is 4, and ndim of input_2 is 3.

The following is allowed in this situation:

`` out = slice_like(a, b, axes=(0, 2))``

Example:

x = [[  1.,   2.,   3.,   4.],
     [  5.,   6.,   7.,   8.],
     [  9.,  10.,  11.,  12.]]

y = [[  0.,   0.,   0.],
     [  0.,   0.,   0.]]

slice_like(x, y) = [[ 1.,  2.,  3.]
                    [ 5.,  6.,  7.]]
slice_like(x, y, axes=(0, 1)) = [[ 1.,  2.,  3.]
                                 [ 5.,  6.,  7.]]
slice_like(x, y, axes=(0)) = [[ 1.,  2.,  3.,  4.]
                              [ 5.,  6.,  7.,  8.]]
slice_like(x, y, axes=(-1)) = [[  1.,   2.,   3.]
                               [  5.,   6.,   7.]
                               [  9.,  10.,  11.]]

Defined in src/operator/tensor/matrix_op.cc:L665

Parameters:
  • data (NDArray) – Source input
  • shape_like (NDArray) – Shape like input
  • axes (Shape(tuple), optional, default=[]) – List of axes on which input data will be sliced according to the corresponding size of the second input. By default will slice on all axes. Negative axes are supported.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.smooth_l1(data=None, scalar=_Null, out=None, name=None, **kwargs)

Calculate Smooth L1 Loss(lhs, scalar) by summing

\[\begin{split}f(x) = \begin{cases} (\sigma x)^2/2,& \text{if }x < 1/\sigma^2\\ |x|-0.5/\sigma^2,& \text{otherwise} \end{cases}\end{split}\]

where \(x\) is an element of the tensor lhs and \(\sigma\) is the scalar.

Example:

smooth_l1([1, 2, 3, 4]) = [0.5, 1.5, 2.5, 3.5]
smooth_l1([1, 2, 3, 4], scalar=1) = [0.5, 1.5, 2.5, 3.5]

Defined in src/operator/tensor/elemwise_binary_scalar_op_extended.cc:L104

Parameters:
  • data (NDArray) – source input
  • scalar (float) – scalar input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.softmax(data=None, axis=_Null, temperature=_Null, dtype=_Null, out=None, name=None, **kwargs)

Applies the softmax function.

The resulting array contains elements in the range (0,1) and the elements along the given axis sum up to 1.

\[softmax(\mathbf{z/t})_j = \frac{e^{z_j/t}}{\sum_{k=1}^K e^{z_k/t}}\]

for \(j = 1, ..., K\)

t is the temperature parameter in softmax function. By default, t equals 1.0

Example:

x = [[ 1.  1.  1.]
     [ 1.  1.  1.]]

softmax(x,axis=0) = [[ 0.5  0.5  0.5]
                     [ 0.5  0.5  0.5]]

softmax(x,axis=1) = [[ 0.33333334,  0.33333334,  0.33333334],
                     [ 0.33333334,  0.33333334,  0.33333334]]

Defined in src/operator/nn/softmax.cc:L93

Parameters:
  • data (NDArray) – The input array.
  • axis (int, optional, default='-1') – The axis along which to compute softmax.
  • temperature (double or None, optional, default=None) – Temperature parameter in softmax
  • dtype ({None, 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to the same as input’s dtype if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.softmax_cross_entropy(data=None, label=None, out=None, name=None, **kwargs)

Calculate cross entropy of softmax output and one-hot label.

  • This operator computes the cross entropy in two steps: - Applies softmax function on the input array. - Computes and returns the cross entropy loss between the softmax output and the labels.

  • The softmax function and cross entropy loss is given by:

    • Softmax Function:
    \[\text{softmax}(x)_i = \frac{exp(x_i)}{\sum_j exp(x_j)}\]
    • Cross Entropy Function:
    \[\text{CE(label, output)} = - \sum_i \text{label}_i \log(\text{output}_i)\]

Example:

x = [[1, 2, 3],
     [11, 7, 5]]

label = [2, 0]

softmax(x) = [[0.09003057, 0.24472848, 0.66524094],
              [0.97962922, 0.01794253, 0.00242826]]

softmax_cross_entropy(data, label) = - log(0.66524084) - log(0.97962922) = 0.4281871

Defined in src/operator/loss_binary_op.cc:L59

Parameters:
  • data (NDArray) – Input data
  • label (NDArray) – Input label
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.softmin(data=None, axis=_Null, temperature=_Null, dtype=_Null, out=None, name=None, **kwargs)

Applies the softmin function.

The resulting array contains elements in the range (0,1) and the elements along the given axis sum up to 1.

\[softmin(\mathbf{z/t})_j = \frac{e^{-z_j/t}}{\sum_{k=1}^K e^{-z_k/t}}\]

for \(j = 1, ..., K\)

t is the temperature parameter in softmax function. By default, t equals 1.0

Example:

x = [[ 1.  2.  3.]
     [ 3.  2.  1.]]

softmin(x,axis=0) = [[ 0.88079703,  0.5,  0.11920292],
                     [ 0.11920292,  0.5,  0.88079703]]

softmin(x,axis=1) = [[ 0.66524094,  0.24472848,  0.09003057],
                     [ 0.09003057,  0.24472848,  0.66524094]]

Defined in src/operator/nn/softmax.cc:L153

Parameters:
  • data (NDArray) – The input array.
  • axis (int, optional, default='-1') – The axis along which to compute softmax.
  • temperature (double or None, optional, default=None) – Temperature parameter in softmax
  • dtype ({None, 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to the same as input’s dtype if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.softsign(data=None, out=None, name=None, **kwargs)

Computes softsign of x element-wise.

\[y = x / (1 + abs(x))\]

The storage type of softsign output is always dense

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L163

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sort(data=None, axis=_Null, is_ascend=_Null, out=None, name=None, **kwargs)

Returns a sorted copy of an input array along the given axis.

Examples:

x = [[ 1, 4],
     [ 3, 1]]

// sorts along the last axis
sort(x) = [[ 1.,  4.],
           [ 1.,  3.]]

// flattens and then sorts
sort(x) = [ 1.,  1.,  3.,  4.]

// sorts along the first axis
sort(x, axis=0) = [[ 1.,  1.],
                   [ 3.,  4.]]

// in a descend order
sort(x, is_ascend=0) = [[ 4.,  1.],
                        [ 3.,  1.]]

Defined in src/operator/tensor/ordering_op.cc:L127

Parameters:
  • data (NDArray) – The input array
  • axis (int or None, optional, default='-1') – Axis along which to choose sort the input tensor. If not given, the flattened array is used. Default is -1.
  • is_ascend (boolean, optional, default=1) – Whether to sort in ascending or descending order.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.space_to_depth(data=None, block_size=_Null, out=None, name=None, **kwargs)

Rearranges(permutes) blocks of spatial data into depth. Similar to ONNX SpaceToDepth operator: https://github.com/onnx/onnx/blob/master/docs/Operators.md#SpaceToDepth

The output is a new tensor where the values from height and width dimension are moved to the depth dimension. The reverse of this operation is depth_to_space.

\[\begin{split}\begin{gather*} x \prime = reshape(x, [N, C, H / block\_size, block\_size, W / block\_size, block\_size]) \\ x \prime \prime = transpose(x \prime, [0, 3, 5, 1, 2, 4]) \\ y = reshape(x \prime \prime, [N, C * (block\_size ^ 2), H / block\_size, W / block\_size]) \end{gather*}\end{split}\]

where \(x\) is an input tensor with default layout as \([N, C, H, W]\): [batch, channels, height, width] and \(y\) is the output tensor of layout \([N, C * (block\_size ^ 2), H / block\_size, W / block\_size]\)

Example:

x = [[[[0, 6, 1, 7, 2, 8],
       [12, 18, 13, 19, 14, 20],
       [3, 9, 4, 10, 5, 11],
       [15, 21, 16, 22, 17, 23]]]]


space_to_depth(x, 2) = [[[[0, 1, 2],
                          [3, 4, 5]],
                         [[6, 7, 8],
                          [9, 10, 11]],
                         [[12, 13, 14],
                          [15, 16, 17]],
                         [[18, 19, 20],
                          [21, 22, 23]]]]

Defined in src/operator/tensor/matrix_op.cc:L1104

Parameters:
  • data (NDArray) – Input ndarray
  • block_size (int, required) – Blocks of [block_size. block_size] are moved
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.split(data=None, num_outputs=_Null, axis=_Null, squeeze_axis=_Null, out=None, name=None, **kwargs)

Splits an array along a particular axis into multiple sub-arrays.

Note

SliceChannel is deprecated. Use split instead.

Note that num_outputs should evenly divide the length of the axis along which to split the array.

Example:

x  = [[[ 1.]
       [ 2.]]
      [[ 3.]
       [ 4.]]
      [[ 5.]
       [ 6.]]]
x.shape = (3, 2, 1)

y = split(x, axis=1, num_outputs=2) // a list of 2 arrays with shape (3, 1, 1)
y = [[[ 1.]]
     [[ 3.]]
     [[ 5.]]]

    [[[ 2.]]
     [[ 4.]]
     [[ 6.]]]

y[0].shape = (3, 1, 1)

z = split(x, axis=0, num_outputs=3) // a list of 3 arrays with shape (1, 2, 1)
z = [[[ 1.]
      [ 2.]]]

    [[[ 3.]
      [ 4.]]]

    [[[ 5.]
      [ 6.]]]

z[0].shape = (1, 2, 1)

squeeze_axis=1 removes the axis with length 1 from the shapes of the output arrays. Note that setting squeeze_axis to 1 removes axis with length 1 only along the axis which it is split. Also squeeze_axis can be set to true only if input.shape[axis] == num_outputs.

Example:

z = split(x, axis=0, num_outputs=3, squeeze_axis=1) // a list of 3 arrays with shape (2, 1)
z = [[ 1.]
     [ 2.]]

    [[ 3.]
     [ 4.]]

    [[ 5.]
     [ 6.]]
z[0].shape = (2 ,1 )

Defined in src/operator/slice_channel.cc:L107

Parameters:
  • data (NDArray) – The input
  • num_outputs (int, required) – Number of splits. Note that this should evenly divide the length of the axis.
  • axis (int, optional, default='1') – Axis along which to split.
  • squeeze_axis (boolean, optional, default=0) – If true, Removes the axis with length 1 from the shapes of the output arrays. Note that setting squeeze_axis to true removes axis with length 1 only along the axis which it is split. Also squeeze_axis can be set to true only if input.shape[axis] == num_outputs.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sqrt(data=None, out=None, name=None, **kwargs)

Returns element-wise square-root value of the input.

\[\textrm{sqrt}(x) = \sqrt{x}\]

Example:

sqrt([4, 9, 16]) = [2, 3, 4]

The storage type of sqrt output depends upon the input storage type:

  • sqrt(default) = default
  • sqrt(row_sparse) = row_sparse
  • sqrt(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L907

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.square(data=None, out=None, name=None, **kwargs)

Returns element-wise squared value of the input.

\[square(x) = x^2\]

Example:

square([2, 3, 4]) = [4, 9, 16]

The storage type of square output depends upon the input storage type:

  • square(default) = default
  • square(row_sparse) = row_sparse
  • square(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L883

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.squeeze(*data, **kwargs)

Remove single-dimensional entries from the shape of an array. Same behavior of defining the output tensor shape as numpy.squeeze for the most of cases. See the following note for exception.

Examples:

data = [[[0], [1], [2]]]
squeeze(data) = [0, 1, 2]
squeeze(data, axis=0) = [[0], [1], [2]]
squeeze(data, axis=2) = [[0, 1, 2]]
squeeze(data, axis=(0, 2)) = [0, 1, 2]

Note

The output of this operator will keep at least one dimension not removed. For example, squeeze([[[4]]]) = [4], while in numpy.squeeze, the output will become a scalar.

Parameters:
  • data (NDArray[]) – data to squeeze
  • axis (Shape or None, optional, default=None) – Selects a subset of the single-dimensional entries in the shape. If an axis is selected with shape entry greater than one, an error is raised.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.stack(*data, **kwargs)

Join a sequence of arrays along a new axis.

The axis parameter specifies the index of the new axis in the dimensions of the result. For example, if axis=0 it will be the first dimension and if axis=-1 it will be the last dimension.

Examples:

x = [1, 2]
y = [3, 4]

stack(x, y) = [[1, 2],
               [3, 4]]
stack(x, y, axis=1) = [[1, 3],
                       [2, 4]]
Parameters:
  • data (NDArray[]) – List of arrays to stack
  • axis (int, optional, default='0') – The axis in the result array along which the input arrays are stacked.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.stop_gradient(data=None, out=None, name=None, **kwargs)

Stops gradient computation.

Stops the accumulated gradient of the inputs from flowing through this operator in the backward direction. In other words, this operator prevents the contribution of its inputs to be taken into account for computing gradients.

Example:

v1 = [1, 2]
v2 = [0, 1]
a = Variable('a')
b = Variable('b')
b_stop_grad = stop_gradient(3 * b)
loss = MakeLoss(b_stop_grad + a)

executor = loss.simple_bind(ctx=cpu(), a=(1,2), b=(1,2))
executor.forward(is_train=True, a=v1, b=v2)
executor.outputs
[ 1.  5.]

executor.backward()
executor.grad_arrays
[ 0.  0.]
[ 1.  1.]

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L299

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sum(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the sum of array elements over given axes.

Note

sum and sum_axis are equivalent. For ndarray of csr storage type summation along axis 0 and axis 1 is supported. Setting keepdims or exclude to True will cause a fallback to dense operator.

Example:

data = [[[1, 2], [2, 3], [1, 3]],
        [[1, 4], [4, 3], [5, 2]],
        [[7, 1], [7, 2], [7, 3]]]

sum(data, axis=1)
[[  4.   8.]
 [ 10.   9.]
 [ 21.   6.]]

sum(data, axis=[1,2])
[ 12.  19.  27.]

data = [[1, 2, 0],
        [3, 0, 1],
        [4, 1, 0]]

csr = cast_storage(data, 'csr')

sum(csr, axis=0)
[ 8.  3.  1.]

sum(csr, axis=1)
[ 3.  4.  5.]

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L116

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.sum_axis(data=None, axis=_Null, keepdims=_Null, exclude=_Null, out=None, name=None, **kwargs)

Computes the sum of array elements over given axes.

Note

sum and sum_axis are equivalent. For ndarray of csr storage type summation along axis 0 and axis 1 is supported. Setting keepdims or exclude to True will cause a fallback to dense operator.

Example:

data = [[[1, 2], [2, 3], [1, 3]],
        [[1, 4], [4, 3], [5, 2]],
        [[7, 1], [7, 2], [7, 3]]]

sum(data, axis=1)
[[  4.   8.]
 [ 10.   9.]
 [ 21.   6.]]

sum(data, axis=[1,2])
[ 12.  19.  27.]

data = [[1, 2, 0],
        [3, 0, 1],
        [4, 1, 0]]

csr = cast_storage(data, 'csr')

sum(csr, axis=0)
[ 8.  3.  1.]

sum(csr, axis=1)
[ 3.  4.  5.]

Defined in src/operator/tensor/broadcast_reduce_op_value.cc:L116

Parameters:
  • data (NDArray) – The input
  • axis (Shape or None, optional, default=None) –

    The axis or axes along which to perform the reduction.

    The default, axis=(), will compute over all elements into a scalar array with shape (1,).

    If axis is int, a reduction is performed on a particular axis.

    If axis is a tuple of ints, a reduction is performed on all the axes specified in the tuple.

    If exclude is true, reduction will be performed on the axes that are NOT in axis instead.

    Negative values means indexing from right to left.

  • keepdims (boolean, optional, default=0) – If this is set to True, the reduced axes are left in the result as dimension with size one.
  • exclude (boolean, optional, default=0) – Whether to perform reduction on axis that are NOT in axis instead.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.swapaxes(data=None, dim1=_Null, dim2=_Null, out=None, name=None, **kwargs)

Interchanges two axes of an array.

Examples:

 x = [[1, 2, 3]])
 swapaxes(x, 0, 1) = [[ 1],
                      [ 2],
                      [ 3]]

 x = [[[ 0, 1],
       [ 2, 3]],
      [[ 4, 5],
       [ 6, 7]]]  // (2,2,2) array

swapaxes(x, 0, 2) = [[[ 0, 4],
                      [ 2, 6]],
                     [[ 1, 5],
                      [ 3, 7]]]

Defined in src/operator/swapaxis.cc:L70

Parameters:
  • data (NDArray) – Input array.
  • dim1 (int (non-negative), optional, default=0) – the first axis to be swapped.
  • dim2 (int (non-negative), optional, default=0) – the second axis to be swapped.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.take(a=None, indices=None, axis=_Null, mode=_Null, out=None, name=None, **kwargs)

Takes elements from an input array along the given axis.

This function slices the input array along a particular axis with the provided indices.

Given data tensor of rank r >= 1, and indices tensor of rank q, gather entries of the axis dimension of data (by default outer-most one as axis=0) indexed by indices, and concatenates them in an output tensor of rank q + (r - 1).

Examples:

x = [4.  5.  6.]

// Trivial case, take the second element along the first axis.

take(x, [1]) = [ 5. ]

// The other trivial case, axis=-1, take the third element along the first axis

take(x, [3], axis=-1, mode='clip') = [ 6. ]

x = [[ 1.,  2.],
     [ 3.,  4.],
     [ 5.,  6.]]

// In this case we will get rows 0 and 1, then 1 and 2. Along axis 0

take(x, [[0,1],[1,2]]) = [[[ 1.,  2.],
                           [ 3.,  4.]],

                          [[ 3.,  4.],
                           [ 5.,  6.]]]

// In this case we will get rows 0 and 1, then 1 and 2 (calculated by wrapping around).
// Along axis 1

take(x, [[0, 3], [-1, -2]], axis=1, mode='wrap') = [[[ 1.  2.]
                                                     [ 2.  1.]]

                                                    [[ 3.  4.]
                                                     [ 4.  3.]]

                                                    [[ 5.  6.]
                                                     [ 6.  5.]]]

The storage type of take output depends upon the input storage type:

  • take(default, default) = default
  • take(csr, default, axis=0) = csr

Defined in src/operator/tensor/indexing_op.cc:L695

Parameters:
  • a (NDArray) – The input array.
  • indices (NDArray) – The indices of the values to be extracted.
  • axis (int, optional, default='0') – The axis of input array to be taken.For input tensor of rank r, it could be in the range of [-r, r-1]
  • mode ({'clip', 'raise', 'wrap'},optional, default='clip') – Specify how out-of-bound indices bahave. Default is “clip”. “clip” means clip to the range. So, if all indices mentioned are too large, they are replaced by the index that addresses the last element along an axis. “wrap” means to wrap around. “raise” means to raise an error, not supported yet.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.tan(data=None, out=None, name=None, **kwargs)

Computes the element-wise tangent of the input array.

The input should be in radians (\(2\pi\) rad equals 360 degrees).

\[tan([0, \pi/4, \pi/2]) = [0, 1, -inf]\]

The storage type of tan output depends upon the input storage type:

  • tan(default) = default
  • tan(row_sparse) = row_sparse
  • tan(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L139

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.tanh(data=None, out=None, name=None, **kwargs)

Returns the hyperbolic tangent of the input array, computed element-wise.

\[tanh(x) = sinh(x) / cosh(x)\]

The storage type of tanh output depends upon the input storage type:

  • tanh(default) = default
  • tanh(row_sparse) = row_sparse
  • tanh(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_trig.cc:L290

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.tile(data=None, reps=_Null, out=None, name=None, **kwargs)

Repeats the whole array multiple times.

If reps has length d, and input array has dimension of n. There are three cases:

  • n=d. Repeat i-th dimension of the input by reps[i] times:

    x = [[1, 2],
         [3, 4]]
    
    tile(x, reps=(2,3)) = [[ 1.,  2.,  1.,  2.,  1.,  2.],
                           [ 3.,  4.,  3.,  4.,  3.,  4.],
                           [ 1.,  2.,  1.,  2.,  1.,  2.],
                           [ 3.,  4.,  3.,  4.,  3.,  4.]]
    
  • n>d. reps is promoted to length n by pre-pending 1’s to it. Thus for an input shape (2,3), repos=(2,) is treated as (1,2):

    tile(x, reps=(2,)) = [[ 1.,  2.,  1.,  2.],
                          [ 3.,  4.,  3.,  4.]]
    
  • n. The input is promoted to be d-dimensional by prepending new axes. So a shape (2,2) array is promoted to (1,2,2) for 3-D replication:

    tile(x, reps=(2,2,3)) = [[[ 1.,  2.,  1.,  2.,  1.,  2.],
                              [ 3.,  4.,  3.,  4.,  3.,  4.],
                              [ 1.,  2.,  1.,  2.,  1.,  2.],
                              [ 3.,  4.,  3.,  4.,  3.,  4.]],
    
                             [[ 1.,  2.,  1.,  2.,  1.,  2.],
                              [ 3.,  4.,  3.,  4.,  3.,  4.],
                              [ 1.,  2.,  1.,  2.,  1.,  2.],
                              [ 3.,  4.,  3.,  4.,  3.,  4.]]]
    

Defined in src/operator/tensor/matrix_op.cc:L857

Parameters:
  • data (NDArray) – Input data array
  • reps (Shape(tuple), required) – The number of times for repeating the tensor a. Each dim size of reps must be a positive integer. If reps has length d, the result will have dimension of max(d, a.ndim); If a.ndim < d, a is promoted to be d-dimensional by prepending new axes. If a.ndim > d, reps is promoted to a.ndim by pre-pending 1’s to it.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.topk(data=None, axis=_Null, k=_Null, ret_typ=_Null, is_ascend=_Null, dtype=_Null, out=None, name=None, **kwargs)
Returns the top k elements in an input array along the given axis.
The returned elements will be sorted.

Examples:

x = [[ 0.3,  0.2,  0.4],
     [ 0.1,  0.3,  0.2]]

// returns an index of the largest element on last axis
topk(x) = [[ 2.],
           [ 1.]]

// returns the value of top-2 largest elements on last axis
topk(x, ret_typ='value', k=2) = [[ 0.4,  0.3],
                                 [ 0.3,  0.2]]

// returns the value of top-2 smallest elements on last axis
topk(x, ret_typ='value', k=2, is_ascend=1) = [[ 0.2 ,  0.3],
                                             [ 0.1 ,  0.2]]

// returns the value of top-2 largest elements on axis 0
topk(x, axis=0, ret_typ='value', k=2) = [[ 0.3,  0.3,  0.4],
                                         [ 0.1,  0.2,  0.2]]

// flattens and then returns list of both values and indices
topk(x, ret_typ='both', k=2) = [[[ 0.4,  0.3], [ 0.3,  0.2]] ,  [[ 2.,  0.], [ 1.,  2.]]]

Defined in src/operator/tensor/ordering_op.cc:L64

Parameters:
  • data (NDArray) – The input array
  • axis (int or None, optional, default='-1') – Axis along which to choose the top k indices. If not given, the flattened array is used. Default is -1.
  • k (int, optional, default='1') – Number of top elements to select, should be always smaller than or equal to the element number in the given axis. A global sort is performed if set k < 1.
  • ret_typ ({'both', 'indices', 'mask', 'value'},optional, default='indices') – The return type. “value” means to return the top k values, “indices” means to return the indices of the top k values, “mask” means to return a mask array containing 0 and 1. 1 means the top k values. “both” means to return a list of both values and indices of top k elements.
  • is_ascend (boolean, optional, default=0) – Whether to choose k largest or k smallest elements. Top K largest elements will be chosen if set to false.
  • dtype ({'float16', 'float32', 'float64', 'int32', 'uint8'},optional, default='float32') – DType of the output indices when ret_typ is “indices” or “both”. An error will be raised if the selected data type cannot precisely represent the indices.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.transpose(data=None, axes=_Null, out=None, name=None, **kwargs)

Permutes the dimensions of an array.

Examples:

x = [[ 1, 2],
     [ 3, 4]]

transpose(x) = [[ 1.,  3.],
                [ 2.,  4.]]

x = [[[ 1.,  2.],
      [ 3.,  4.]],

     [[ 5.,  6.],
      [ 7.,  8.]]]

transpose(x) = [[[ 1.,  5.],
                 [ 3.,  7.]],

                [[ 2.,  6.],
                 [ 4.,  8.]]]

transpose(x, axes=(1,0,2)) = [[[ 1.,  2.],
                               [ 5.,  6.]],

                              [[ 3.,  4.],
                               [ 7.,  8.]]]

Defined in src/operator/tensor/matrix_op.cc:L375

Parameters:
  • data (NDArray) – Source input
  • axes (Shape(tuple), optional, default=[]) – Target axis order. By default the axes will be inverted.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.trunc(data=None, out=None, name=None, **kwargs)

Return the element-wise truncated value of the input.

The truncated value of the scalar x is the nearest integer i which is closer to zero than x is. In short, the fractional part of the signed number x is discarded.

Example:

trunc([-2.1, -1.9, 1.5, 1.9, 2.1]) = [-2., -1.,  1.,  1.,  2.]

The storage type of trunc output depends upon the input storage type:

  • trunc(default) = default
  • trunc(row_sparse) = row_sparse
  • trunc(csr) = csr

Defined in src/operator/tensor/elemwise_unary_op_basic.cc:L825

Parameters:
  • data (NDArray) – The input array.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.uniform(low=_Null, high=_Null, shape=_Null, ctx=_Null, dtype=_Null, out=None, name=None, **kwargs)

Draw random samples from a uniform distribution.

Note

The existing alias uniform is deprecated.

Samples are uniformly distributed over the half-open interval [low, high) (includes low, but excludes high).

Example:

uniform(low=0, high=1, shape=(2,2)) = [[ 0.60276335,  0.85794562],
                                       [ 0.54488319,  0.84725171]]

Defined in src/operator/random/sample_op.cc:L96

Parameters:
  • low (float, optional, default=0) – Lower bound of the distribution.
  • high (float, optional, default=1) – Upper bound of the distribution.
  • shape (Shape(tuple), optional, default=None) – Shape of the output.
  • ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n). Only used for imperative calls.
  • dtype ({'None', 'float16', 'float32', 'float64'},optional, default='None') – DType of the output in case this can’t be inferred. Defaults to float32 if not defined (dtype=None).
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.unravel_index(data=None, shape=_Null, out=None, name=None, **kwargs)

Converts an array of flat indices into a batch of index arrays. The operator follows numpy conventions so a single multi index is given by a column of the output matrix. The leading dimension may be left unspecified by using -1 as placeholder.

Examples:

A = [22,41,37]
unravel(A, shape=(7,6)) = [[3,6,6],[4,5,1]]
unravel(A, shape=(-1,6)) = [[3,6,6],[4,5,1]]

Defined in src/operator/tensor/ravel.cc:L67

Parameters:
  • data (NDArray) – Array of flat indices
  • shape (Shape(tuple), optional, default=None) – Shape of the array into which the multi-indices apply.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.where(condition=None, x=None, y=None, out=None, name=None, **kwargs)

Return the elements, either from x or y, depending on the condition.

Given three ndarrays, condition, x, and y, return an ndarray with the elements from x or y, depending on the elements from condition are true or false. x and y must have the same shape. If condition has the same shape as x, each element in the output array is from x if the corresponding element in the condition is true, and from y if false.

If condition does not have the same shape as x, it must be a 1D array whose size is the same as x’s first dimension size. Each row of the output array is from x’s row if the corresponding element from condition is true, and from y’s row if false.

Note that all non-zero values are interpreted as True in condition.

Examples:

x = [[1, 2], [3, 4]]
y = [[5, 6], [7, 8]]
cond = [[0, 1], [-1, 0]]

where(cond, x, y) = [[5, 2], [3, 8]]

csr_cond = cast_storage(cond, 'csr')

where(csr_cond, x, y) = [[5, 2], [3, 8]]

Defined in src/operator/tensor/control_flow_op.cc:L57

Parameters:
  • condition (NDArray) – condition array
  • x (NDArray) –
  • y (NDArray) –
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.zeros_like(data=None, out=None, name=None, **kwargs)

Return an array of zeros with the same shape, type and storage type as the input array.

The storage type of zeros_like output depends on the storage type of the input

  • zeros_like(row_sparse) = row_sparse
  • zeros_like(csr) = csr
  • zeros_like(default) = default

Examples:

x = [[ 1.,  1.,  1.],
     [ 1.,  1.,  1.]]

zeros_like(x) = [[ 0.,  0.,  0.],
                 [ 0.,  0.,  0.]]
Parameters:
  • data (NDArray) – The input
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays

mxnet.ndarray.concatenate(arrays, axis=0, always_copy=True)[source]

DEPRECATED, use concat instead

Parameters:
  • arrays (list of NDArray) – Arrays to be concatenate. They must have identical shape except the first dimension. They also must have the same data type.
  • axis (int) – The axis along which to concatenate.
  • always_copy (bool) – Default True. When not True, if the arrays only contain one NDArray, that element will be returned directly, avoid copying.
Returns:

An NDArray that lives on the same context as arrays[0].context.

Return type:

NDArray

mxnet.ndarray.ones(shape, ctx=None, dtype=None, **kwargs)[source]

Returns a new array filled with all ones, with the given shape and type.

Parameters:
  • shape (int or tuple of int or list of int) – The shape of the empty array.
  • ctx (Context, optional) – An optional device context. Defaults to the current default context (mxnet.context.current_context()).
  • dtype (str or numpy.dtype, optional) – An optional value type (default is float32).
  • out (NDArray, optional) – The output NDArray (default is None).
Returns:

A new array of the specified shape filled with all ones.

Return type:

NDArray

Examples

>>> mx.nd.ones(1).asnumpy()
array([ 1.], dtype=float32)
>>> mx.nd.ones((1,2), mx.gpu(0))

>>> mx.nd.ones((1,2), dtype='float16').asnumpy()
array([[ 1.,  1.]], dtype=float16)
mxnet.ndarray.add(lhs, rhs)[source]

Returns element-wise sum of the input arrays with broadcasting.

Equivalent to lhs + rhs, mx.nd.broadcast_add(lhs, rhs) and mx.nd.broadcast_plus(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be added.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be added. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

The element-wise sum of the input arrays.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x+2).asnumpy()
array([[ 3.,  3.,  3.],
       [ 3.,  3.,  3.]], dtype=float32)
>>> (x+y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 2.,  2.,  2.]], dtype=float32)
>>> mx.nd.add(x,y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 2.,  2.,  2.]], dtype=float32)
>>> (z + y).asnumpy()
array([[ 0.,  1.],
       [ 1.,  2.]], dtype=float32)
mxnet.ndarray.arange(start, stop=None, step=1.0, repeat=1, infer_range=None, ctx=None, dtype=)[source]

Returns evenly spaced values within a given interval.

Values are generated within the half-open interval [start, stop). In other words, the interval includes start but excludes stop. The function is similar to the built-in Python function range and to numpy.arange, but returns an NDArray.

Parameters:
  • start (number, optional) – Start of interval. The default start value is 0.
  • stop (number) – End of interval.
  • step (number, optional) – Spacing between values. The default step size is 1.
  • repeat (int, optional) – Number of times to repeat each element. The default repeat count is 1.
  • infer_range (boolean, optional) – When set to True, infer the stop position from the start, step, repeat, and output tensor size.
  • ctx (Context, optional) – Device context. Default context is the current default context.
  • dtype (str or numpy.dtype, optional) – The data type of the NDArray. The default datatype is np.float32.
Returns:

NDArray of evenly spaced values in the specified range.

Return type:

NDArray

Examples

>>> mx.nd.arange(3).asnumpy()
array([ 0.,  1.,  2.], dtype=float32)
>>> mx.nd.arange(2, 6).asnumpy()
array([ 2.,  3.,  4.,  5.], dtype=float32)
>>> mx.nd.arange(2, 6, step=2).asnumpy()
array([ 2.,  4.], dtype=float32)
>>> mx.nd.arange(2, 6, step=1.5, repeat=2).asnumpy()
array([ 2. ,  2. ,  3.5,  3.5,  5. ,  5. ], dtype=float32)
>>> mx.nd.arange(2, 6, step=2, repeat=3, dtype='int32').asnumpy()
array([2, 2, 2, 4, 4, 4], dtype=int32)
mxnet.ndarray.linspace(start, stop, num, endpoint=True, ctx=None, dtype=)[source]

Return evenly spaced numbers within a specified interval.

Values are generated within the half-open interval [start, stop) or closed interval [start, stop] depending on whether endpoint is True or False. The function is similar to numpy.linspace, but returns an NDArray.

Parameters:
  • start (number) – Start of interval.
  • stop (number) – End of interval, unless endpoint is set to False. In that case, the sequence consists of all but the last of num + 1 evenly spaced samples, so that stop is excluded. Note that the step size changes when endpoint is False.
  • num (number) – Number of samples to generate. Must be non-negative.
  • endpoint (bool) – If True, stop is the last sample. Otherwise, it is not included. The default is True.
  • ctx (Context, optional) – Device context. Default context is the current default context.
  • dtype (str or numpy.dtype, optional) – The data type of the NDArray. The default datatype is np.float32.
Returns:

NDArray of evenly spaced values in the specified range.

Return type:

NDArray

Examples

>>> mx.nd.linspace(2.0, 3.0, 5).asnumpy()
array([ 2.,  2.25.,  2.5,  2.75,  3.], dtype=float32)
>>> mx.nd.linspace(2.0, 3.0, 5, endpoint=False).asnumpy()
array([ 2.,  2.2.,  2.4,  2.6,  2.8], dtype=float32)
mxnet.ndarray.eye(N, M=0, k=0, ctx=None, dtype=None, **kwargs)[source]

Return a 2-D array with ones on the diagonal and zeros elsewhere.

Parameters:
  • N (int) – Number of rows in the output.
  • M (int, optional) – Number of columns in the output. If 0, defaults to N.
  • k (int, optional) – Index of the diagonal: 0 (the default) refers to the main diagonal, a positive value refers to an upper diagonal, and a negative value to a lower diagonal.
  • ctx (Context, optional) – An optional device context (default is the current default context)
  • dtype (str or numpy.dtype, optional) – An optional value type (default is float32)
Returns:

A created array

Return type:

NDArray

Examples

>>> mx.nd.eye(2)
[[ 1.  0.]
 [ 0.  1.]]

>>> mx.nd.eye(2, 3, 1)
[[ 0.  1.  0.]
 [ 0.  0.  1.]]

mxnet.ndarray.divide(lhs, rhs)[source]

Returns element-wise division of the input arrays with broadcasting.

Equivalent to lhs / rhs and mx.nd.broadcast_div(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array in division.
  • rhs (scalar or mxnet.ndarray.array) – Second array in division. The arrays to be divided. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

The element-wise division of the input arrays.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))*6
>>> y = mx.nd.ones((2,1))*2
>>> x.asnumpy()
array([[ 6.,  6.,  6.],
       [ 6.,  6.,  6.]], dtype=float32)
>>> y.asnumpy()
array([[ 2.],
       [ 2.]], dtype=float32)
>>> x/2

>>> (x/3).asnumpy()
array([[ 2.,  2.,  2.],
       [ 2.,  2.,  2.]], dtype=float32)
>>> (x/y).asnumpy()
array([[ 3.,  3.,  3.],
       [ 3.,  3.,  3.]], dtype=float32)
>>> mx.nd.divide(x,y).asnumpy()
array([[ 3.,  3.,  3.],
       [ 3.,  3.,  3.]], dtype=float32)
mxnet.ndarray.equal(lhs, rhs)[source]

Returns the result of element-wise equal to (==) comparison operation with broadcasting.

For each element in input arrays, return 1(true) if corresponding elements are same, otherwise return 0(false).

Equivalent to lhs == rhs and mx.nd.broadcast_equal(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x == 1).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (x == y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.equal(x,y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (z == y).asnumpy()
array([[ 1.,  0.],
       [ 0.,  1.]], dtype=float32)
mxnet.ndarray.full(shape, val, ctx=None, dtype=, out=None)[source]

Returns a new array of given shape and type, filled with the given value val.

Parameters:
  • shape (int or tuple of int) – The shape of the new array.
  • val (scalar) – Fill value.
  • ctx (Context, optional) – Device context (default is the current default context).
  • dtype (str or numpy.dtype, optional) – The data type of the returned NDArray. The default datatype is float32.
  • out (NDArray, optional) – The output NDArray (default is None).
Returns:

NDArray filled with val, with the given shape, ctx, and dtype.

Return type:

NDArray

Examples

>>> mx.nd.full(1, 2.0).asnumpy()
array([ 2.], dtype=float32)
>>> mx.nd.full((1, 2), 2.0, mx.gpu(0))

>>> mx.nd.full((1, 2), 2.0, dtype='float16').asnumpy()
array([[ 2.,  2.]], dtype=float16)
mxnet.ndarray.greater(lhs, rhs)[source]

Returns the result of element-wise greater than (>) comparison operation with broadcasting.

For each element in input arrays, return 1(true) if lhs elements are greater than rhs, otherwise return 0(false).

Equivalent to lhs > rhs and mx.nd.broadcast_greater(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x > 1).asnumpy()
array([[ 0.,  0.,  0.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> (x > y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> mx.nd.greater(x, y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> (z > y).asnumpy()
array([[ 0.,  1.],
       [ 0.,  0.]], dtype=float32)
mxnet.ndarray.greater_equal(lhs, rhs)[source]

Returns the result of element-wise greater than or equal to (>=) comparison operation with broadcasting.

For each element in input arrays, return 1(true) if lhs elements are greater than equal to rhs, otherwise return 0(false).

Equivalent to lhs >= rhs and mx.nd.broadcast_greater_equal(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x >= 1).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (x >= y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.greater_equal(x, y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (z >= y).asnumpy()
array([[ 1.,  1.],
       [ 0.,  1.]], dtype=float32)
mxnet.ndarray.imdecode(str_img, clip_rect=(0, 0, 0, 0), out=None, index=0, channels=3, mean=None)[source]

DEPRECATED, use mx.img instead

Parameters:
  • str_img (str) – Binary image data
  • clip_rect (iterable of 4 int) – Clip decoded image to rectangle (x0, y0, x1, y1).
  • out (NDArray) – Output buffer. Can be 3 dimensional (c, h, w) or 4 dimensional (n, c, h, w).
  • index (int) – Output decoded image to i-th slice of 4 dimensional buffer.
  • channels (int) – Number of channels to output. Decode to grey scale when channels = 1.
  • mean (NDArray) – Subtract mean from decode image before outputing.
mxnet.ndarray.lesser(lhs, rhs)[source]

Returns the result of element-wise lesser than (<) comparison operation with broadcasting.

For each element in input arrays, return 1(true) if lhs elements are less than rhs, otherwise return 0(false).

Equivalent to lhs < rhs and mx.nd.broadcast_lesser(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x < 1).asnumpy()
array([[ 0.,  0.,  0.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> (x < y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> mx.nd.lesser(x, y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> (z < y).asnumpy()
array([[ 0.,  0.],
       [ 1.,  0.]], dtype=float32)
mxnet.ndarray.lesser_equal(lhs, rhs)[source]

Returns the result of element-wise lesser than or equal to (<=) comparison operation with broadcasting.

For each element in input arrays, return 1(true) if lhs elements are lesser than equal to rhs, otherwise return 0(false).

Equivalent to lhs <= rhs and mx.nd.broadcast_lesser_equal(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x <= 1).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (x <= y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.lesser_equal(x, y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (z <= y).asnumpy()
array([[ 1.,  0.],
       [ 1.,  1.]], dtype=float32)
mxnet.ndarray.logical_and(lhs, rhs)[source]

Returns the result of element-wise logical and comparison operation with broadcasting.

For each element in input arrays, return 1(true) if lhs elements and rhs elements are true, otherwise return 0(false).

Equivalent to lhs and rhs and mx.nd.broadcast_logical_and(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First input of the function.
  • rhs (scalar or mxnet.ndarray.array) – Second input of the function. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> mx.nd.logical_and(x, 1).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.logical_and(x, y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.logical_and(z, y).asnumpy()
array([[ 0.,  0.],
       [ 0.,  1.]], dtype=float32)
mxnet.ndarray.logical_or(lhs, rhs)[source]

Returns the result of element-wise logical or comparison operation with broadcasting.

For each element in input arrays, return 1(true) if lhs elements or rhs elements are true, otherwise return 0(false).

Equivalent to lhs or rhs and mx.nd.broadcast_logical_or(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First input of the function.
  • rhs (scalar or mxnet.ndarray.array) – Second input of the function. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> mx.nd.logical_or(x, 1).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.logical_or(x, y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.logical_or(z, y).asnumpy()
array([[ 0.,  1.],
       [ 1.,  1.]], dtype=float32)
mxnet.ndarray.logical_xor(lhs, rhs)[source]

Returns the result of element-wise logical xor comparison operation with broadcasting.

For each element in input arrays, return 1(true) if lhs elements or rhs elements are true, otherwise return 0(false).

Equivalent to bool(lhs) ^ bool(rhs) and mx.nd.broadcast_logical_xor(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First input of the function.
  • rhs (scalar or mxnet.ndarray.array) – Second input of the function. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> mx.nd.logical_xor(x, y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 0.,  0.,  0.]], dtype=float32)
mxnet.ndarray.maximum(lhs, rhs)[source]

Returns element-wise maximum of the input arrays with broadcasting.

Equivalent to mx.nd.broadcast_maximum(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

The element-wise maximum of the input arrays.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> mx.nd.maximum(x, 2).asnumpy()
array([[ 2.,  2.,  2.],
       [ 2.,  2.,  2.]], dtype=float32)
>>> mx.nd.maximum(x, y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.maximum(y, z).asnumpy()
array([[ 0.,  1.],
       [ 1.,  1.]], dtype=float32)
mxnet.ndarray.minimum(lhs, rhs)[source]

Returns element-wise minimum of the input arrays with broadcasting.

Equivalent to mx.nd.broadcast_minimum(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

The element-wise minimum of the input arrays.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> mx.nd.minimum(x, 2).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.minimum(x, y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.minimum(z, y).asnumpy()
array([[ 0.,  0.],
       [ 0.,  1.]], dtype=float32)
mxnet.ndarray.moveaxis(tensor, source, destination)[source]

Moves the source axis into the destination position while leaving the other axes in their original order

Parameters:
  • tensor (mx.nd.array) – The array which axes should be reordered
  • source (int or sequence of int) – Original position of the axes to move. Can be negative but must be unique.
  • destination (int or sequence of int) – Destination position for each of the original axes. Can be negative but must be unique.
Returns:

result – Array with moved axes.

Return type:

mx.nd.array

Examples

>>> X = mx.nd.array([[1, 2, 3], [4, 5, 6]])
>>> mx.nd.moveaxis(X, 0, 1).shape
(3L, 2L)
>>> X = mx.nd.zeros((3, 4, 5))
>>> mx.nd.moveaxis(X, [0, 1], [-1, -2]).shape
(5, 4, 3)
mxnet.ndarray.modulo(lhs, rhs)[source]

Returns element-wise modulo of the input arrays with broadcasting.

Equivalent to lhs % rhs and mx.nd.broadcast_mod(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array in modulo.
  • rhs (scalar or mxnet.ndarray.array) – Second array in modulo. The arrays to be taken modulo. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

The element-wise modulo of the input arrays.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))*6
>>> y = mx.nd.ones((2,1))*4
>>> x.asnumpy()
array([[ 6.,  6.,  6.],
       [ 6.,  6.,  6.]], dtype=float32)
>>> y.asnumpy()
array([[ 4.],
       [ 4.]], dtype=float32)
>>> x%5

>>> (x%5).asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (x%y).asnumpy()
array([[ 2.,  2.,  2.],
       [ 2.,  2.,  2.]], dtype=float32)
>>> mx.nd.modulo(x,y).asnumpy()
array([[ 2.,  2.,  2.],
       [ 2.,  2.,  2.]], dtype=float32)
mxnet.ndarray.multiply(lhs, rhs)[source]

Returns element-wise product of the input arrays with broadcasting.

Equivalent to lhs * rhs and mx.nd.broadcast_mul(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be multiplied.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be multiplied. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

The element-wise multiplication of the input arrays.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x*2).asnumpy()
array([[ 2.,  2.,  2.],
       [ 2.,  2.,  2.]], dtype=float32)
>>> (x*y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> mx.nd.multiply(x, y).asnumpy()
array([[ 0.,  0.,  0.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> (z*y).asnumpy()
array([[ 0.,  0.],
       [ 0.,  1.]], dtype=float32)
mxnet.ndarray.not_equal(lhs, rhs)[source]

Returns the result of element-wise not equal to (!=) comparison operation with broadcasting.

For each element in input arrays, return 1(true) if corresponding elements are different, otherwise return 0(false).

Equivalent to lhs != rhs and mx.nd.broadcast_not_equal(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be compared.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be compared. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

Output array of boolean values.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (z == y).asnumpy()
array([[ 1.,  0.],
       [ 0.,  1.]], dtype=float32)
>>> (x != 1).asnumpy()
array([[ 0.,  0.,  0.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> (x != y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> mx.nd.not_equal(x, y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> (z != y).asnumpy()
array([[ 0.,  1.],
       [ 1.,  0.]], dtype=float32)
mxnet.ndarray.onehot_encode(indices, out)[source]

One-hot encoding indices into matrix out.

Note

onehot_encode is deprecated. Use one_hot instead.

mxnet.ndarray.power(base, exp)[source]

Returns result of first array elements raised to powers from second array, element-wise with broadcasting.

Equivalent to base ** exp and mx.nd.broadcast_power(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • base (scalar or NDArray) – The base array
  • exp (scalar or NDArray) – The exponent array. If base.shape != exp.shape, they must be broadcastable to a common shape.
Returns:

The bases in x raised to the exponents in y.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))*2
>>> y = mx.nd.arange(1,3).reshape((2,1))
>>> z = mx.nd.arange(1,3).reshape((2,1))
>>> x.asnumpy()
array([[ 2.,  2.,  2.],
       [ 2.,  2.,  2.]], dtype=float32)
>>> y.asnumpy()
array([[ 1.],
       [ 2.]], dtype=float32)
>>> z.asnumpy()
array([[ 1.],
       [ 2.]], dtype=float32)
>>> (x**2).asnumpy()
array([[ 4.,  4.,  4.],
       [ 4.,  4.,  4.]], dtype=float32)
>>> (x**y).asnumpy()
array([[ 2.,  2.,  2.],
       [ 4.,  4.,  4.]], dtype=float32)
>>> mx.nd.power(x,y).asnumpy()
array([[ 2.,  2.,  2.],
       [ 4.,  4.,  4.]], dtype=float32)
>>> (z**y).asnumpy()
array([[ 1.],
       [ 4.]], dtype=float32)
mxnet.ndarray.subtract(lhs, rhs)[source]

Returns element-wise difference of the input arrays with broadcasting.

Equivalent to lhs - rhs, mx.nd.broadcast_sub(lhs, rhs) and mx.nd.broadcast_minus(lhs, rhs).

Note

If the corresponding dimensions of two arrays have the same size or one of them has size 1, then the arrays are broadcastable to a common shape.

Parameters:
  • lhs (scalar or mxnet.ndarray.array) – First array to be subtracted.
  • rhs (scalar or mxnet.ndarray.array) – Second array to be subtracted. If lhs.shape != rhs.shape, they must be broadcastable to a common shape.
Returns:

The element-wise difference of the input arrays.

Return type:

NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.arange(2).reshape((2,1))
>>> z = mx.nd.arange(2).reshape((1,2))
>>> x.asnumpy()
array([[ 1.,  1.,  1.],
       [ 1.,  1.,  1.]], dtype=float32)
>>> y.asnumpy()
array([[ 0.],
       [ 1.]], dtype=float32)
>>> z.asnumpy()
array([[ 0.,  1.]], dtype=float32)
>>> (x-2).asnumpy()
array([[-1., -1., -1.],
       [-1., -1., -1.]], dtype=float32)
>>> (x-y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> mx.nd.subtract(x,y).asnumpy()
array([[ 1.,  1.,  1.],
       [ 0.,  0.,  0.]], dtype=float32)
>>> (z-y).asnumpy()
array([[ 0.,  1.],
       [-1.,  0.]], dtype=float32)
mxnet.ndarray.true_divide(lhs, rhs)[source]

This function is similar to divide().

mxnet.ndarray.waitall()[source]

Wait for all async operations to finish in MXNet.

This function is used for benchmarking only.

Note

If your mxnet code throws an exception, then waitall can cause performance impact.

mxnet.ndarray.histogram(a, bins=10, range=None)[source]

Compute the histogram of the input data.

Parameters:
  • a (NDArray) – Input data. The histogram is computed over the flattened array.
  • bins (int or sequence of scalars) – If bins is an int, it defines the number of equal-width bins in the given range (10, by default). If bins is a sequence, it defines the bin edges, including the rightmost edge, allowing for non-uniform bin widths.
  • range ((float, float), optional) – The lower and upper range of the bins. If not provided, range is simply (a.min(), a.max()). Values outside the range are ignored. The first element of the range must be less than or equal to the second. range affects the automatic bin computation as well, the range will be equally divided by the number of bins.
Returns:

A created array.

Return type:

NDArray

mxnet.ndarray.split_v2(ary, indices_or_sections, axis=0, squeeze_axis=False)[source]

Split an array into multiple sub-arrays.

Parameters:
  • ary (NDArray) – Array to be divided into sub-arrays.
  • indices_or_sections (int or tuple of ints) – If indices_or_sections is an integer, N, the array will be divided into N equal arrays along axis. If such a split is not possible, an error is raised. If indices_or_sections is a 1-D array of sorted integers, the entries indicate where along axis the array is split. For example, [2, 3] would, for axis=0, result in - ary[:2] - ary[2:3] - ary[3:] If an index exceeds the dimension of the array along axis, an empty sub-array is returned correspondingly.
  • axis (int, optional) – The axis along which to split, default is 0.
  • squeeze_axis (boolean, optional) – Whether to squeeze the axis of sub-arrays or not, only useful when size of the sub-arrays are 1 on the axis. Default is False.
Returns:

A created array.

Return type:

NDArray

mxnet.ndarray.to_dlpack_for_read(data)[source]
Returns a reference view of NDArray that represents as DLManagedTensor until
all previous write operations on the current array are finished.
Parameters:data (NDArray) – input data.
Returns:a reference view of NDArray that represents as DLManagedTensor.
Return type:PyCapsule (the pointer of DLManagedTensor)

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.to_dlpack_for_read(x)
>>> type(y)

>>> z = mx.nd.from_dlpack(y)
>>> z
[[1. 1. 1.]
 [1. 1. 1.]]

mxnet.ndarray.to_dlpack_for_write(data)[source]
Returns a reference view of NDArray that represents as DLManagedTensor until
all previous read/write operations on the current array are finished.
Parameters:data (NDArray) – input data.
Returns:a reference view of NDArray that represents as DLManagedTensor.
Return type:PyCapsule (the pointer of DLManagedTensor)

Examples

>>> x = mx.nd.ones((2,3))
>>> w = mx.nd.to_dlpack_for_write(x)
>>> type(w)

>>> u = mx.nd.from_dlpack(w)
>>> u += 1
>>> x
[[2. 2. 2.]
 [2. 2. 2.]]

mxnet.ndarray.from_dlpack(dlpack)[source]

Returns a NDArray backed by a dlpack tensor.

Parameters:dlpack (PyCapsule (the pointer of DLManagedTensor)) – input data
Returns:a NDArray backed by a dlpack tensor
Return type:NDArray

Examples

>>> x = mx.nd.ones((2,3))
>>> y = mx.nd.to_dlpack_for_read(x)
>>> type(y)

>>> z = mx.nd.from_dlpack(y)
>>> type(z)

>>> z
[[ 1.  1.  1.]
 [ 1.  1.  1.]]

>>> w = mx.nd.to_dlpack_for_write(x)
>>> type(w)

>>> u = mx.nd.from_dlpack(w)
>>> u += 1
>>> x
[[2. 2. 2.]
 [2. 2. 2.]]

mxnet.ndarray.from_numpy(ndarray, zero_copy=True)[source]

Returns an MXNet’s ndarray backed by numpy’s ndarray. When zero_copy is set to be true, this API consumes numpy’s ndarray and produces MXNet’s ndarray without having to copy the content. In this case, we disallow users to modify the given numpy ndarray, and it is suggested not to read the numpy ndarray as well for internal correctness.

Parameters:
  • ndarray (numpy.ndarray) – input data
  • zero_copy (bool) – Whether we use DLPack’s zero-copy conversion to convert to MXNet’s NDArray. This is only available for c-contiguous arrays, i.e. array.flags[C_CONTIGUOUS] == True.
Returns:

a NDArray backed by a dlpack tensor

Return type:

NDArray

mxnet.ndarray.zeros(shape, ctx=None, dtype=None, stype=None, **kwargs)[source]

Return a new array of given shape and type, filled with zeros.

Parameters:
  • shape (int or tuple of int) – The shape of the empty array
  • ctx (Context, optional) – An optional device context (default is the current default context)
  • dtype (str or numpy.dtype, optional) – An optional value type (default is float32)
  • stype (string, optional) – The storage type of the empty array, such as ‘row_sparse’, ‘csr’, etc.
Returns:

A created array

Return type:

NDArray, CSRNDArray or RowSparseNDArray

Examples

>>> mx.nd.zeros((1,2), mx.cpu(), stype='csr')

>>> mx.nd.zeros((1,2), mx.cpu(), 'float16', stype='row_sparse').asnumpy()
array([[ 0.,  0.]], dtype=float16)
mxnet.ndarray.empty(shape, ctx=None, dtype=None, stype=None)[source]

Returns a new array of given shape and type, without initializing entries.

Parameters:
  • shape (int or tuple of int) – The shape of the empty array.
  • ctx (Context, optional) – An optional device context (default is the current default context).
  • dtype (str or numpy.dtype, optional) – An optional value type (default is float32).
  • stype (str, optional) – An optional storage type (default is default).
Returns:

A created array.

Return type:

NDArray, CSRNDArray or RowSparseNDArray

Examples

>>> mx.nd.empty(1)

>>> mx.nd.empty((1,2), mx.gpu(0))

>>> mx.nd.empty((1,2), mx.gpu(0), 'float16')

>>> mx.nd.empty((1,2), stype='csr')

mxnet.ndarray.array(source_array, ctx=None, dtype=None)[source]

Creates an array from any object exposing the array interface.

Parameters:
  • source_array (array_like) – An object exposing the array interface, an object whose __array__ method returns an array, or any (nested) sequence.
  • ctx (Context, optional) – Device context (default is the current default context).
  • dtype (str or numpy.dtype, optional) – The data type of the output array. The default dtype is source_array.dtype if source_array is an NDArray, float32 otherwise.
Returns:

An array with the same contents as the source_array.

Return type:

NDArray, RowSparseNDArray or CSRNDArray

Examples

>>> import numpy as np
>>> mx.nd.array([1, 2, 3])

>>> mx.nd.array([[1, 2], [3, 4]])

>>> mx.nd.array(np.zeros((3, 2)))

>>> mx.nd.array(np.zeros((3, 2)), mx.gpu(0))

>>> mx.nd.array(mx.nd.zeros((3, 2), stype='row_sparse'))

mxnet.ndarray.load_frombuffer(buf)[source]

Loads an array dictionary or list from a buffer

See more details in save.

Parameters:buf (str) – Buffer containing contents of a file as a string or bytes.
Returns:Loaded data.
Return type:list of NDArray, RowSparseNDArray or CSRNDArray, or dict of str to NDArray, RowSparseNDArray or CSRNDArray

NDArray API of MXNet.

mxnet.ndarray.save(fname, data)[source]

Saves a list of arrays or a dict of str->array to file.

Examples of filenames:

  • /path/to/file
  • s3://my-bucket/path/to/file (if compiled with AWS S3 supports)
  • hdfs://path/to/file (if compiled with HDFS supports)
Parameters:

Examples

>>> x = mx.nd.zeros((2,3))
>>> y = mx.nd.ones((1,4))
>>> mx.nd.save('my_list', [x,y])
>>> mx.nd.save('my_dict', {'x':x, 'y':y})
>>> mx.nd.load('my_list')
[, ]
>>> mx.nd.load('my_dict')
{'y': , 'x': }
mxnet.ndarray.load(fname)[source]

Loads an array from file.

See more details in save.

Parameters:fname (str) – The filename.
Returns:Loaded data.
Return type:list of NDArray, RowSparseNDArray or CSRNDArray, or dict of str to NDArray, RowSparseNDArray or CSRNDArray

Random number interface of MXNet.

mxnet.random.seed(seed_state, ctx='all')[source]

Seeds the random number generators in MXNet.

This affects the behavior of modules in MXNet that uses random number generators, like the dropout operator and NDArray‘s random sampling operators.

Parameters:
  • seed_state (int) – The random number seed.
  • ctx (Context) – The device context of the generator. The default is “all” which means seeding random number generators of all devices.

Notes

Random number generators in MXNet are device specific. mx.random.seed(seed_state) sets the state of each generator using seed_state and the device id. Therefore, random numbers generated from different devices can be different even if they are seeded using the same seed.

To produce identical random number sequences independent of the device id, set optional ctx argument. This produces the same sequence of random numbers independent of the device id, but the sequence can be different on different kind of devices as MXNet’s random number generators for CPU and GPU use different algorithms.

Example

>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 1.36481571 -0.62203991]
 [-1.4962182  -0.08511394]]
>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 1.09544981 -0.20014545]
 [-0.20808885  0.2527658 ]]
# Same results on the same device with the same seed
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 0.47400656 -0.75213492]
 [ 0.20251541  0.95352972]]
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2)).asnumpy())
[[ 0.47400656 -0.75213492]
 [ 0.20251541  0.95352972]]
# Different results on gpu(0) and gpu(1) with the same seed
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(0)).asnumpy())
[[ 2.5020072 -1.6884501]
 [-0.7931333 -1.4218881]]
>>> mx.random.seed(128)
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(1)).asnumpy())
[[ 0.24336822 -1.664805  ]
 [-1.0223296   1.253198  ]]
# Seeding with `ctx` argument produces identical results on gpu(0) and gpu(1)
>>> mx.random.seed(128, ctx=mx.gpu(0))
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(0)).asnumpy())
[[ 2.5020072 -1.6884501]
 [-0.7931333 -1.4218881]]
>>> mx.random.seed(128, ctx=mx.gpu(1))
>>> print(mx.nd.random.normal(shape=(2,2), ctx=mx.gpu(1)).asnumpy())
[[ 2.5020072 -1.6884501]
 [-0.7931333 -1.4218881]]