# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements. See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership. The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
# coding: utf-8
"""Definition of various recurrent neural network cells."""
__all__ = ['VariationalDropoutCell', 'LSTMPCell']
from ...rnn import BidirectionalCell, SequentialRNNCell, ModifierCell, HybridRecurrentCell
from ...rnn.rnn_cell import _format_sequence, _get_begin_state, _mask_sequence_variable_length
from ... import tensor_types
from ....base import _as_list
[docs]class VariationalDropoutCell(ModifierCell):
"""
Applies Variational Dropout on base cell.
https://arxiv.org/pdf/1512.05287.pdf
Variational dropout uses the same dropout mask across time-steps. It can be applied to RNN
inputs, outputs, and states. The masks for them are not shared.
The dropout mask is initialized when stepping forward for the first time and will remain
the same until .reset() is called. Thus, if using the cell and stepping manually without calling
.unroll(), the .reset() should be called after each sequence.
Parameters
----------
base_cell : RecurrentCell
The cell on which to perform variational dropout.
drop_inputs : float, default 0.
The dropout rate for inputs. Won't apply dropout if it equals 0.
drop_states : float, default 0.
The dropout rate for state inputs on the first state channel.
Won't apply dropout if it equals 0.
drop_outputs : float, default 0.
The dropout rate for outputs. Won't apply dropout if it equals 0.
"""
def __init__(self, base_cell, drop_inputs=0., drop_states=0., drop_outputs=0.):
assert not drop_states or not isinstance(base_cell, BidirectionalCell), \
"BidirectionalCell doesn't support variational state dropout. " \
"Please add VariationalDropoutCell to the cells underneath instead."
assert not drop_states \
or not isinstance(base_cell, SequentialRNNCell) or not base_cell._bidirectional, \
"Bidirectional SequentialRNNCell doesn't support variational state dropout. " \
"Please add VariationalDropoutCell to the cells underneath instead."
super(VariationalDropoutCell, self).__init__(base_cell)
self.drop_inputs = drop_inputs
self.drop_states = drop_states
self.drop_outputs = drop_outputs
self.drop_inputs_mask = None
self.drop_states_mask = None
self.drop_outputs_mask = None
def _alias(self):
return 'vardrop'
def reset(self):
super(VariationalDropoutCell, self).reset()
self.drop_inputs_mask = None
self.drop_states_mask = None
self.drop_outputs_mask = None
def _initialize_input_masks(self, F, inputs, states):
if self.drop_states and self.drop_states_mask is None:
self.drop_states_mask = F.Dropout(F.ones_like(states[0]),
p=self.drop_states)
if self.drop_inputs and self.drop_inputs_mask is None:
self.drop_inputs_mask = F.Dropout(F.ones_like(inputs),
p=self.drop_inputs)
def _initialize_output_mask(self, F, output):
if self.drop_outputs and self.drop_outputs_mask is None:
self.drop_outputs_mask = F.Dropout(F.ones_like(output),
p=self.drop_outputs)
def hybrid_forward(self, F, inputs, states):
cell = self.base_cell
self._initialize_input_masks(F, inputs, states)
if self.drop_states:
states = list(states)
# state dropout only needs to be applied on h, which is always the first state.
states[0] = states[0] * self.drop_states_mask
if self.drop_inputs:
inputs = inputs * self.drop_inputs_mask
next_output, next_states = cell(inputs, states)
self._initialize_output_mask(F, next_output)
if self.drop_outputs:
next_output = next_output * self.drop_outputs_mask
return next_output, next_states
def __repr__(self):
s = '{name}(p_out = {drop_outputs}, p_state = {drop_states})'
return s.format(name=self.__class__.__name__,
**self.__dict__)
[docs] def unroll(self, length, inputs, begin_state=None, layout='NTC', merge_outputs=None,
valid_length=None):
"""Unrolls an RNN cell across time steps.
Parameters
----------
length : int
Number of steps to unroll.
inputs : Symbol, list of Symbol, or None
If `inputs` is a single Symbol (usually the output
of Embedding symbol), it should have shape
(batch_size, length, ...) if `layout` is 'NTC',
or (length, batch_size, ...) if `layout` is 'TNC'.
If `inputs` is a list of symbols (usually output of
previous unroll), they should all have shape
(batch_size, ...).
begin_state : nested list of Symbol, optional
Input states created by `begin_state()`
or output state of another cell.
Created from `begin_state()` if `None`.
layout : str, optional
`layout` of input symbol. Only used if inputs
is a single Symbol.
merge_outputs : bool, optional
If `False`, returns outputs as a list of Symbols.
If `True`, concatenates output across time steps
and returns a single symbol with shape
(batch_size, length, ...) if layout is 'NTC',
or (length, batch_size, ...) if layout is 'TNC'.
If `None`, output whatever is faster.
valid_length : Symbol, NDArray or None
`valid_length` specifies the length of the sequences in the batch without padding.
This option is especially useful for building sequence-to-sequence models where
the input and output sequences would potentially be padded.
If `valid_length` is None, all sequences are assumed to have the same length.
If `valid_length` is a Symbol or NDArray, it should have shape (batch_size,).
The ith element will be the length of the ith sequence in the batch.
The last valid state will be return and the padded outputs will be masked with 0.
Note that `valid_length` must be smaller or equal to `length`.
Returns
-------
outputs : list of Symbol or Symbol
Symbol (if `merge_outputs` is True) or list of Symbols
(if `merge_outputs` is False) corresponding to the output from
the RNN from this unrolling.
states : list of Symbol
The new state of this RNN after this unrolling.
The type of this symbol is same as the output of `begin_state()`.
"""
# Dropout on inputs and outputs can be performed on the whole sequence
# only when state dropout is not present.
if self.drop_states:
return super(VariationalDropoutCell, self).unroll(length, inputs, begin_state,
layout, merge_outputs,
valid_length=valid_length)
self.reset()
inputs, axis, F, batch_size = _format_sequence(length, inputs, layout, True)
states = _get_begin_state(self, F, begin_state, inputs, batch_size)
if self.drop_inputs:
inputs = F.Dropout(inputs, p=self.drop_inputs, axes=(axis,))
outputs, states = self.base_cell.unroll(length, inputs, states, layout, merge_outputs=True,
valid_length=valid_length)
if self.drop_outputs:
outputs = F.Dropout(outputs, p=self.drop_outputs, axes=(axis,))
merge_outputs = isinstance(outputs, tensor_types) if merge_outputs is None else \
merge_outputs
outputs, _, _, _ = _format_sequence(length, outputs, layout, merge_outputs)
if valid_length is not None:
outputs = _mask_sequence_variable_length(F, outputs, length, valid_length, axis,
merge_outputs)
return outputs, states
[docs]class LSTMPCell(HybridRecurrentCell):
r"""Long-Short Term Memory Projected (LSTMP) network cell.
(https://arxiv.org/abs/1402.1128)
Each call computes the following function:
.. math::
\begin{array}{ll}
i_t = sigmoid(W_{ii} x_t + b_{ii} + W_{ri} r_{(t-1)} + b_{ri}) \\
f_t = sigmoid(W_{if} x_t + b_{if} + W_{rf} r_{(t-1)} + b_{rf}) \\
g_t = \tanh(W_{ig} x_t + b_{ig} + W_{rc} r_{(t-1)} + b_{rg}) \\
o_t = sigmoid(W_{io} x_t + b_{io} + W_{ro} r_{(t-1)} + b_{ro}) \\
c_t = f_t * c_{(t-1)} + i_t * g_t \\
h_t = o_t * \tanh(c_t) \\
r_t = W_{hr} h_t
\end{array}
where :math:`r_t` is the projected recurrent activation at time `t`,
:math:`h_t` is the hidden state at time `t`, :math:`c_t` is the
cell state at time `t`, :math:`x_t` is the input at time `t`, and :math:`i_t`,
:math:`f_t`, :math:`g_t`, :math:`o_t` are the input, forget, cell, and
out gates, respectively.
Parameters
----------
hidden_size : int
Number of units in cell state symbol.
projection_size : int
Number of units in output symbol.
i2h_weight_initializer : str or Initializer
Initializer for the input weights matrix, used for the linear
transformation of the inputs.
h2h_weight_initializer : str or Initializer
Initializer for the recurrent weights matrix, used for the linear
transformation of the hidden state.
h2r_weight_initializer : str or Initializer
Initializer for the projection weights matrix, used for the linear
transformation of the recurrent state.
i2h_bias_initializer : str or Initializer, default 'lstmbias'
Initializer for the bias vector. By default, bias for the forget
gate is initialized to 1 while all other biases are initialized
to zero.
h2h_bias_initializer : str or Initializer
Initializer for the bias vector.
prefix : str, default ``'lstmp_``'
Prefix for name of `Block`s
(and name of weight if params is `None`).
params : Parameter or None
Container for weight sharing between cells.
Created if `None`.
Inputs:
- **data**: input tensor with shape `(batch_size, input_size)`.
- **states**: a list of two initial recurrent state tensors, with shape
`(batch_size, projection_size)` and `(batch_size, hidden_size)` respectively.
Outputs:
- **out**: output tensor with shape `(batch_size, num_hidden)`.
- **next_states**: a list of two output recurrent state tensors. Each has
the same shape as `states`.
"""
def __init__(self, hidden_size, projection_size,
i2h_weight_initializer=None, h2h_weight_initializer=None,
h2r_weight_initializer=None,
i2h_bias_initializer='zeros', h2h_bias_initializer='zeros',
input_size=0, prefix=None, params=None):
super(LSTMPCell, self).__init__(prefix=prefix, params=params)
self._hidden_size = hidden_size
self._input_size = input_size
self._projection_size = projection_size
self.i2h_weight = self.params.get('i2h_weight', shape=(4*hidden_size, input_size),
init=i2h_weight_initializer,
allow_deferred_init=True)
self.h2h_weight = self.params.get('h2h_weight', shape=(4*hidden_size, projection_size),
init=h2h_weight_initializer,
allow_deferred_init=True)
self.h2r_weight = self.params.get('h2r_weight', shape=(projection_size, hidden_size),
init=h2r_weight_initializer,
allow_deferred_init=True)
self.i2h_bias = self.params.get('i2h_bias', shape=(4*hidden_size,),
init=i2h_bias_initializer,
allow_deferred_init=True)
self.h2h_bias = self.params.get('h2h_bias', shape=(4*hidden_size,),
init=h2h_bias_initializer,
allow_deferred_init=True)
def state_info(self, batch_size=0):
return [{'shape': (batch_size, self._projection_size), '__layout__': 'NC'},
{'shape': (batch_size, self._hidden_size), '__layout__': 'NC'}]
def _alias(self):
return 'lstmp'
def __repr__(self):
s = '{name}({mapping})'
shape = self.i2h_weight.shape
proj_shape = self.h2r_weight.shape
mapping = '{0} -> {1} -> {2}'.format(shape[1] if shape[1] else None,
shape[0], proj_shape[0])
return s.format(name=self.__class__.__name__,
mapping=mapping,
**self.__dict__)
# pylint: disable= arguments-differ
def hybrid_forward(self, F, inputs, states, i2h_weight,
h2h_weight, h2r_weight, i2h_bias, h2h_bias):
prefix = 't%d_'%self._counter
i2h = F.FullyConnected(data=inputs, weight=i2h_weight, bias=i2h_bias,
num_hidden=self._hidden_size*4, name=prefix+'i2h')
h2h = F.FullyConnected(data=states[0], weight=h2h_weight, bias=h2h_bias,
num_hidden=self._hidden_size*4, name=prefix+'h2h')
gates = i2h + h2h
slice_gates = F.SliceChannel(gates, num_outputs=4, name=prefix+'slice')
in_gate = F.Activation(slice_gates[0], act_type="sigmoid", name=prefix+'i')
forget_gate = F.Activation(slice_gates[1], act_type="sigmoid", name=prefix+'f')
in_transform = F.Activation(slice_gates[2], act_type="tanh", name=prefix+'c')
out_gate = F.Activation(slice_gates[3], act_type="sigmoid", name=prefix+'o')
next_c = F.elemwise_add(forget_gate * states[1], in_gate * in_transform,
name=prefix+'state')
hidden = F.elemwise_mul(out_gate, F.Activation(next_c, act_type="tanh"),
name=prefix+'hidden')
next_r = F.FullyConnected(data=hidden, num_hidden=self._projection_size,
weight=h2r_weight, no_bias=True, name=prefix+'out')
return next_r, [next_r, next_c]
# pylint: enable= arguments-differ
def dynamic_unroll(cell, inputs, begin_state, drop_inputs=0, drop_outputs=0,
layout='TNC', valid_length=None):
"""Unrolls an RNN cell across time steps.
Currently, 'TNC' is a preferred layout. unroll on the input of this layout
runs much faster.
Parameters
----------
cell : an object whose base class is RNNCell.
The RNN cell to run on the input sequence.
inputs : Symbol
It should have shape (batch_size, length, ...) if `layout` is 'NTC',
or (length, batch_size, ...) if `layout` is 'TNC'.
begin_state : nested list of Symbol
The initial states of the RNN sequence.
drop_inputs : float, default 0.
The dropout rate for inputs. Won't apply dropout if it equals 0.
drop_outputs : float, default 0.
The dropout rate for outputs. Won't apply dropout if it equals 0.
layout : str, optional
`layout` of input symbol. Only used if inputs
is a single Symbol.
valid_length : Symbol, NDArray or None
`valid_length` specifies the length of the sequences in the batch without padding.
This option is especially useful for building sequence-to-sequence models where
the input and output sequences would potentially be padded.
If `valid_length` is None, all sequences are assumed to have the same length.
If `valid_length` is a Symbol or NDArray, it should have shape (batch_size,).
The ith element will be the length of the ith sequence in the batch.
The last valid state will be return and the padded outputs will be masked with 0.
Note that `valid_length` must be smaller or equal to `length`.
Returns
-------
outputs : Symbol
the output of the RNN from this unrolling.
states : list of Symbol
The new state of this RNN after this unrolling.
The type of this symbol is same as the output of `begin_state`.
Examples
--------
>>> seq_len = 3
>>> batch_size = 2
>>> input_size = 5
>>> cell = mx.gluon.rnn.LSTMCell(input_size, prefix='rnn_')
>>> cell.initialize(ctx=mx.cpu())
>>> rnn_data = mx.nd.normal(loc=0, scale=1, shape=(seq_len, batch_size, input_size))
>>> state_shape = (batch_size, input_size)
>>> states = [mx.nd.normal(loc=0, scale=1, shape=state_shape) for i in range(2)]
>>> valid_length = mx.nd.array([2, 3])
>>> output, states = mx.gluon.contrib.rnn.rnn_cell.dynamic_unroll(cell, rnn_data, states,
valid_length=valid_length,
layout='TNC')
>>> print(output)
[[[ 0.00767238 0.00023103 0.03973929 -0.00925503 -0.05660512]
[ 0.00881535 0.05428379 -0.02493718 -0.01834097 0.02189514]]
[[-0.00676967 0.01447039 0.01287002 -0.00574152 -0.05734247]
[ 0.01568508 0.02650866 -0.04270559 -0.04328435 0.00904011]]
[[ 0. 0. 0. 0. 0. ]
[ 0.01055336 0.02734251 -0.03153727 -0.03742751 -0.01378113]]]
"""
# Merge is always True, so we don't need length.
inputs, axis, F, _ = _format_sequence(0, inputs, layout, True)
if axis != 0:
axes = list(range(len(layout)))
tmp = axes[0]
axes[0] = axes[axis]
axes[axis] = tmp
inputs = F.transpose(inputs, axes=axes)
states = begin_state
if drop_inputs:
inputs = F.Dropout(inputs, p=drop_inputs, axes=(axis,))
if valid_length is None:
def loop_body(inputs, states):
return cell(inputs, states)
else:
zeros = []
for s in states:
zeros.append(F.zeros_like(s))
states = list(_as_list(states))
states.append(F.zeros((1)))
def loop_body(inputs, states):
cell_states = states[:-1]
iter_no = states[-1]
out, new_states = cell(inputs, cell_states)
for i, state in enumerate(cell_states):
new_states[i] = F.where(F.broadcast_greater(valid_length, iter_no),
new_states[i], state)
new_states.append(iter_no + 1)
return out, new_states
outputs, states = F.contrib.foreach(loop_body, inputs, states)
if drop_outputs:
outputs = F.Dropout(outputs, p=drop_outputs, axes=(axis,))
if valid_length is not None:
if axis != 0:
outputs = F.transpose(outputs, axes)
outputs = F.SequenceMask(outputs, sequence_length=valid_length,
use_sequence_length=True, axis=axis)
# the last state is the iteration number. We don't need it.
return outputs, states[:-1]
else:
if axis != 0:
outputs = F.transpose(outputs, axes)
return outputs, states
