Hybrid - Faster training and easy deployment

Related Content:

Deep learning frameworks can be roughly divided into two categories: declarative and imperative. With declarative frameworks (including Tensorflow, Theano, etc) users first declare a fixed computation graph and then execute it end-to-end. The benefit of fixed computation graph is it’s portable and runs more efficiently. However, it’s less flexible because any logic must be encoded into the graph as special operators like scan, while_loop and cond. It’s also hard to debug.

Imperative frameworks (including PyTorch, Chainer, etc) are just the opposite: they execute commands one-by-one just like old fashioned Matlab and Numpy. This style is more flexible, easier to debug, but less efficient.

HybridBlock seamlessly combines declarative programming and imperative programming to offer the benefit of both. Users can quickly develop and debug models with imperative programming and switch to efficient declarative execution by simply calling: HybridBlock.hybridize().

HybridBlock

HybridBlock is very similar to Block but has a few restrictions:

  • All children layers of HybridBlock must also be HybridBlock.
  • Only methods that are implemented for both NDArray and Symbol can be used. For example you cannot use .asnumpy(), .shape, etc.
  • Operations cannot change from run to run. For example, you cannot do if x: if x is different for each iteration.

To use hybrid support, we subclass the HybridBlock:

import mxnet as mx
from mxnet import gluon
from mxnet.gluon import nn

mx.random.seed(42)

class Net(gluon.HybridBlock):
    def __init__(self, **kwargs):
        super(Net, self).__init__(**kwargs)
        with self.name_scope():
            # layers created in name_scope will inherit name space
            # from parent layer.
            self.conv1 = nn.Conv2D(6, kernel_size=5)
            self.pool1 = nn.MaxPool2D(pool_size=2)
            self.conv2 = nn.Conv2D(16, kernel_size=5)
            self.pool2 = nn.MaxPool2D(pool_size=2)
            self.fc1 = nn.Dense(120)
            self.fc2 = nn.Dense(84)
            # You can use a Dense layer for fc3 but we do dot product manually
            # here for illustration purposes.
            self.fc3_weight = self.params.get('fc3_weight', shape=(10, 84))

    def hybrid_forward(self, F, x, fc3_weight):
        # Here `F` can be either mx.nd or mx.sym, x is the input data,
        # and fc3_weight is either self.fc3_weight.data() or
        # self.fc3_weight.var() depending on whether x is Symbol or NDArray
        print(x)
        x = self.pool1(F.relu(self.conv1(x)))
        x = self.pool2(F.relu(self.conv2(x)))
        # 0 means copy over size from corresponding dimension.
        # -1 means infer size from the rest of dimensions.
        x = x.reshape((0, -1))
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = F.dot(x, fc3_weight, transpose_b=True)
        return x

Hybridize

By default, HybridBlock runs just like a standard Block. Each time a layer is called, its hybrid_forward will be run:

net = Net()
net.initialize()
x = mx.nd.random_normal(shape=(16, 1, 28, 28))
net(x)
x = mx.nd.random_normal(shape=(16, 1, 28, 28))
net(x)

Hybrid execution can be activated by simply calling .hybridize() on the top level layer. The first forward call after activation will try to build a computation graph from hybrid_forward and cache it. On subsequent forward calls the cached graph, instead of hybrid_forward, will be invoked:

net.hybridize()
x = mx.nd.random_normal(shape=(16, 1, 28, 28))
net(x)
x = mx.nd.random_normal(shape=(16, 1, 28, 28))
net(x)

Note that before hybridize, print(x) printed out one NDArray for forward, but after hybridize, only the first forward printed out a Symbol. On subsequent forward hybrid_forward is not called so nothing was printed.

Hybridize will speed up execution and save memory. If the top level layer is not a HybridBlock, you can still call .hybridize() on it and Gluon will try to hybridize its children layers instead.

hybridize also accepts several options for performance tuning. For example, you can do

net.hybridize(static_alloc=True)
# or
net.hybridize(static_alloc=True, static_shape=True)

Please refer to the API manual for details.

Serializing trained model for deployment

Models implemented as HybridBlock can be easily serialized. The serialized model can be loaded back later or used for deployment with other language front-ends like C, C++ and Scala. To this end, we simply use export and SymbolBlock.imports:

net(x)
net.export('model', epoch=1)

Two files model-symbol.json and model-0001.params are saved on disk. You can use other language bindings to load them. You can also load them back to gluon with SymbolBlock:

import warnings

with warnings.catch_warnings():
    warnings.simplefilter("ignore")
    net2 = gluon.SymbolBlock.imports('model-symbol.json', ['data'], 'model-0001.params')

Operators that do not work with hybridize

If you want to hybridize your model, you must use F.some_operator in your ‘hybrid_forward’ function. F will be mxnet.nd before you hybridize and mxnet.sym after hybridize. While most APIs are the same in NDArray and Symbol, there are some differences. Writing F.some_operator and call hybridize may not work all of the time. Here we list some frequently used NDArray APIs that can’t be hybridized and provide you the work arounds.

Element-wise Operators

In NDArray APIs, the following arithmetic and comparison APIs are automatically broadcasted if the input NDArrays have different shapes. However, that’s not the case in Symbol API. It’s not automatically broadcasted, and you have to manually specify to use another set of broadcast operators for Symbols expected to have different shapes.

NDArray APIs Description
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._*mul_* x._mul_(y) <=> x*y <=> mx.nd.multiply(x, y)
NDArray._*div_* x._div_(y) <=> x/y <=> mx.nd.divide(x, y)
NDArray._*mod_* x._mod_(y) <=> x%y <=> mx.nd.modulo(x, y)
NDArray._*lt_* x._lt_(y) <=> x 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)

The current workaround is to use corresponding broadcast operators for arithmetic and comparison to avoid potential hybridization failure when input shapes are different.

Symbol APIs Description
broadcast_add Returns element-wise sum of the input arrays with broadcasting.
broadcast_sub Returns element-wise difference of the input arrays with broadcasting.
broadcast_mul Returns element-wise product of the input arrays with broadcasting.
broadcast_div Returns element-wise division of the input arrays with broadcasting.
broadcast_mod Returns element-wise modulo of the input arrays with broadcasting.
broadcast_equal Returns the result of element-wise equal to (==) comparison operation with broadcasting.
broadcast_not_equal Returns the result of element-wise not equal to (!=) comparison operation with broadcasting.
broadcast_greater Returns the result of element-wise greater than (>) comparison operation with broadcasting.
broadcast_greater_equal Returns the result of element-wise greater than or equal to (>=) comparison operation with broadcasting.
broadcast_lesser :: Returns the result of element-wise lesser than (<) comparison operation with broadcasting.
broadcast_lesser_equal Returns the result of element-wise lesser than or equal to (<=) comparison operation with broadcasting.

For example, if you want to add a NDarray to your input x, use broadcast_add instead of +:

def hybrid_forward(self, F, x):
    # avoid writing: return x + F.ones((1, 1))
    return F.broadcast_add(x, F.ones((1, 1)))

If you used +, it would still work before hybridization, but will throw an error of shape missmtach after hybridization.

Shape

Gluon’s imperative interface is very flexible and allows you to print the shape of the NDArray. However, Symbol does not have shape attributes. As a result, you need to avoid printing shapes in hybrid_forward. Otherwise, you will get the following error:

AttributeError: 'Symbol' object has no attribute 'shape'

Slice

[] in NDArray is used to get a slice from the array. However, [] in Symbol is used to get an output from a grouped symbol. For example, you will get different results for the following method before and after hybridization.

def hybrid_forward(self, F, x):
    return x[0]

The current workaround is to explicitly call slice or slice_axis operators in hybrid_forward.

Not implemented operators

Some of the often used operators in NDArray are not implemented in Symbol, and will cause hybridization failure.

NDArray.asnumpy

Symbol does not support the asnumpy function. You need to avoid calling asnumpy in hybrid_forward.

Array creation APIs

mx.nd.array() is used a lot, but Symbol does not have the array API. The current workaround is to use F.ones, F.zeros, or F.full, which exist in both the NDArray and Symbol APIs.

In-Place Arithmetic Operators

In-place arithmetic operators may be used in Gluon imperative mode, however if you expect to hybridize, you should write these operations explicitly instead. For example, avoid writing x += y and use x = x + y, otherwise you will get NotImplementedError. This applies to all the following operators:

NDArray in-place arithmetic operators Description
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

Summary

The recommended practice is to utilize the flexibility of imperative NDArray API during experimentation. Once you finalized your model, make necessary changes mentioned above so you can call hybridize function to improve performance.