def forward(self, x):
"""Compute the stft transform.
Args:
x (Variable): shape(B, T), dtype flaot32, the input waveform.
Returns:
(real, imag)
real (Variable): shape(B, C, 1, T), dtype flaot32, the real part of the spectrogram. (C = 1 + n_fft // 2)
imag (Variable): shape(B, C, 1, T), dtype flaot32, the image part of the spectrogram. (C = 1 + n_fft // 2)
"""
# x(batch_size, time_steps)
# pad it first with reflect mode
pad_start = F.reverse(x[:, 1:1 + self.n_fft // 2], axis=1)
pad_stop = F.reverse(x[:, -(1 + self.n_fft // 2):-1], axis=1)
x = F.concat([pad_start, x, pad_stop], axis=-1)
# to BC1T, C=1
x = F.unsqueeze(x, axes=[1, 2])
out = conv2d(x, self.weight, stride=(1, self.hop_length))
real, imag = F.split(out, 2, dim=1) # BC1T
return real, imag
def basic_lstm(input,
init_hidden,
init_cell,
hidden_size,
num_layers=1,
sequence_length=None,
dropout_prob=0.0,
bidirectional=False,
batch_first=True,
param_attr=None,
bias_attr=None,
gate_activation=None,
activation=None,
forget_bias=1.0,
dtype='float32',
name='basic_lstm'):
"""
LSTM implementation using basic operators, supports multiple layers and bidirectional LSTM.
.. math::
i_t &= \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + b_i)
f_t &= \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + b_f + forget_bias )
o_t &= \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + b_o)
\\tilde{c_t} &= tanh(W_{cx}x_t + W_{ch}h_{t-1} + b_c)
c_t &= f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
h_t &= o_t \odot tanh(c_t)
Args:
input (Variable): lstm input tensor,
if batch_first = False, shape should be ( seq_len x batch_size x input_size )
if batch_first = True, shape should be ( batch_size x seq_len x hidden_size )
init_hidden(Variable|None): The initial hidden state of the LSTM
This is a tensor with shape ( num_layers x batch_size x hidden_size)
if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
and can be reshaped to a tensor with shape ( num_layers x 2 x batch_size x hidden_size) to use.
If it's None, it will be set to all 0.
init_cell(Variable|None): The initial hidden state of the LSTM
This is a tensor with shape ( num_layers x batch_size x hidden_size)
if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
and can be reshaped to a tensor with shape ( num_layers x 2 x batch_size x hidden_size) to use.
If it's None, it will be set to all 0.
hidden_size (int): Hidden size of the LSTM
num_layers (int): The total number of layers of the LSTM
sequence_length (Variabe|None): A tensor (shape [batch_size]) stores each real length of each instance,
This tensor will be convert to a mask to mask the padding ids
If it's None means NO padding ids
dropout_prob(float|0.0): Dropout prob, dropout ONLY work after rnn output of each layers,
NOT between time steps
bidirectional (bool|False): If it is bidirectional
batch_first (bool|True): The shape format of the input and output tensors. If true,
the shape format should be :attr:`[batch_size, seq_len, hidden_size]`. If false,
the shape format should be :attr:`[seq_len, batch_size, hidden_size]`. By default
this function accepts input and emits output in batch-major form to be consistent
with most of data format, though a bit less efficient because of extra transposes.
param_attr(ParamAttr|None): The parameter attribute for the learnable
weight matrix. Note:
If it is set to None or one attribute of ParamAttr, lstm_unit will
create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|None): The parameter attribute for the bias
of LSTM unit.
If it is set to None or one attribute of ParamAttr, lstm_unit will
create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
gate_activation (function|None): The activation function for gates (actGate).
Default: 'fluid.layers.sigmoid'
activation (function|None): The activation function for cell (actNode).
Default: 'fluid.layers.tanh'
forget_bias (float|1.0) : Forget bias used to compute the forget gate
dtype(string): Data type used in this unit
name(string): Name used to identify parameters and biases
Returns:
rnn_out(Tensor), last_hidden(Tensor), last_cell(Tensor)
- rnn_out is the result of LSTM hidden, shape is (seq_len x batch_size x hidden_size) \
if is_bidirec set to True, it's shape will be ( seq_len x batch_sze x hidden_size*2)
- last_hidden is the hidden state of the last step of LSTM \
with shape ( num_layers x batch_size x hidden_size ) \
if is_bidirec set to True, it's shape will be ( num_layers*2 x batch_size x hidden_size),
and can be reshaped to a tensor ( num_layers x 2 x batch_size x hidden_size) to use.
- last_cell is the hidden state of the last step of LSTM \
with shape ( num_layers x batch_size x hidden_size ) \
if is_bidirec set to True, it's shape will be ( num_layers*2 x batch_size x hidden_size),
and can be reshaped to a tensor ( num_layers x 2 x batch_size x hidden_size) to use.
Examples:
.. code-block:: python
import paddle.fluid.layers as layers
from paddle.fluid.contrib.layers import basic_lstm
batch_size = 20
input_size = 128
hidden_size = 256
num_layers = 2
dropout = 0.5
bidirectional = True
batch_first = False
input = layers.data( name = "input", shape = [-1, batch_size, input_size], dtype='float32')
pre_hidden = layers.data( name = "pre_hidden", shape=[-1, hidden_size], dtype='float32')
pre_cell = layers.data( name = "pre_cell", shape=[-1, hidden_size], dtype='float32')
sequence_length = layers.data( name="sequence_length", shape=[-1], dtype='int32')
rnn_out, last_hidden, last_cell = basic_lstm( input, pre_hidden, pre_cell, \
hidden_size, num_layers = num_layers, \
sequence_length = sequence_length, dropout_prob=dropout, bidirectional = bidirectional, \
batch_first = batch_first)
"""
fw_unit_list = []
for i in range(num_layers):
new_name = name + "_layers_" + str(i)
if param_attr is not None and param_attr.name is not None:
layer_param_attr = copy.deepcopy(param_attr)
layer_param_attr.name += "_fw_w_" + str(i)
else:
layer_param_attr = param_attr
if bias_attr is not None and bias_attr.name is not None:
layer_bias_attr = copy.deepcopy(bias_attr)
layer_bias_attr.name += "_fw_b_" + str(i)
else:
layer_bias_attr = bias_attr
fw_unit_list.append(
BasicLSTMUnit(new_name,
hidden_size,
param_attr=layer_param_attr,
bias_attr=layer_bias_attr,
gate_activation=gate_activation,
activation=activation,
forget_bias=forget_bias,
dtype=dtype))
if bidirectional:
bw_unit_list = []
for i in range(num_layers):
new_name = name + "_reverse_layers_" + str(i)
if param_attr is not None and param_attr.name is not None:
layer_param_attr = copy.deepcopy(param_attr)
layer_param_attr.name += "_bw_w_" + str(i)
else:
layer_param_attr = param_attr
if bias_attr is not None and bias_attr.name is not None:
layer_bias_attr = copy.deepcopy(bias_attr)
layer_bias_attr.name += "_bw_b_" + str(i)
else:
layer_bias_attr = param_attr
bw_unit_list.append(
BasicLSTMUnit(new_name,
hidden_size,
param_attr=layer_param_attr,
bias_attr=layer_bias_attr,
gate_activation=gate_activation,
activation=activation,
forget_bias=forget_bias,
dtype=dtype))
if batch_first:
input = layers.transpose(input, [1, 0, 2])
mask = None
if sequence_length:
max_seq_len = layers.shape(input)[0]
mask = layers.sequence_mask(sequence_length,
maxlen=max_seq_len,
dtype='float32')
mask = layers.transpose(mask, [1, 0])
direc_num = 1
if bidirectional:
direc_num = 2
# convert to [num_layers, 2, batch_size, hidden_size]
if init_hidden:
init_hidden = layers.reshape(
init_hidden, shape=[num_layers, direc_num, -1, hidden_size])
init_cell = layers.reshape(
init_cell, shape=[num_layers, direc_num, -1, hidden_size])
# forward direction
def get_single_direction_output(rnn_input,
unit_list,
mask=None,
direc_index=0):
rnn = StaticRNN()
with rnn.step():
step_input = rnn.step_input(rnn_input)
if mask:
step_mask = rnn.step_input(mask)
for i in range(num_layers):
if init_hidden:
pre_hidden = rnn.memory(init=init_hidden[i, direc_index])
pre_cell = rnn.memory(init=init_cell[i, direc_index])
else:
pre_hidden = rnn.memory(batch_ref=rnn_input,
shape=[-1, hidden_size])
pre_cell = rnn.memory(batch_ref=rnn_input,
shape=[-1, hidden_size])
new_hidden, new_cell = unit_list[i](step_input, pre_hidden,
pre_cell)
if mask:
new_hidden = layers.elementwise_mul(
new_hidden, step_mask,
axis=0) - layers.elementwise_mul(pre_hidden,
(step_mask - 1),
axis=0)
new_cell = layers.elementwise_mul(
new_cell, step_mask, axis=0) - layers.elementwise_mul(
pre_cell, (step_mask - 1), axis=0)
rnn.update_memory(pre_hidden, new_hidden)
rnn.update_memory(pre_cell, new_cell)
rnn.step_output(new_hidden)
rnn.step_output(new_cell)
step_input = new_hidden
if dropout_prob != None and dropout_prob > 0.0:
step_input = layers.dropout(
step_input,
dropout_prob=dropout_prob,
dropout_implementation='upscale_in_train')
rnn.step_output(step_input)
rnn_out = rnn()
last_hidden_array = []
last_cell_array = []
rnn_output = rnn_out[-1]
for i in range(num_layers):
last_hidden = rnn_out[i * 2]
last_hidden = last_hidden[-1]
last_hidden_array.append(last_hidden)
last_cell = rnn_out[i * 2 + 1]
last_cell = last_cell[-1]
last_cell_array.append(last_cell)
last_hidden_output = layers.concat(last_hidden_array, axis=0)
last_hidden_output = layers.reshape(
last_hidden_output, shape=[num_layers, -1, hidden_size])
last_cell_output = layers.concat(last_cell_array, axis=0)
last_cell_output = layers.reshape(last_cell_output,
shape=[num_layers, -1, hidden_size])
return rnn_output, last_hidden_output, last_cell_output
# seq_len, batch_size, hidden_size
fw_rnn_out, fw_last_hidden, fw_last_cell = get_single_direction_output(
input, fw_unit_list, mask, direc_index=0)
if bidirectional:
bw_input = layers.reverse(input, axis=[0])
bw_mask = None
if mask:
bw_mask = layers.reverse(mask, axis=[0])
bw_rnn_out, bw_last_hidden, bw_last_cell = get_single_direction_output(
bw_input, bw_unit_list, bw_mask, direc_index=1)
bw_rnn_out = layers.reverse(bw_rnn_out, axis=[0])
rnn_out = layers.concat([fw_rnn_out, bw_rnn_out], axis=2)
last_hidden = layers.concat([fw_last_hidden, bw_last_hidden], axis=1)
last_hidden = layers.reshape(
last_hidden, shape=[num_layers * direc_num, -1, hidden_size])
last_cell = layers.concat([fw_last_cell, bw_last_cell], axis=1)
last_cell = layers.reshape(
last_cell, shape=[num_layers * direc_num, -1, hidden_size])
if batch_first:
rnn_out = layers.transpose(rnn_out, [1, 0, 2])
return rnn_out, last_hidden, last_cell
else:
rnn_out = fw_rnn_out
last_hidden = fw_last_hidden
last_cell = fw_last_cell
if batch_first:
rnn_out = layers.transpose(rnn_out, [1, 0, 2])
return rnn_out, last_hidden, last_cell
def basic_gru(input,
init_hidden,
hidden_size,
num_layers=1,
sequence_length=None,
dropout_prob=0.0,
bidirectional=False,
batch_first=True,
param_attr=None,
bias_attr=None,
gate_activation=None,
activation=None,
dtype='float32',
name='basic_gru'):
"""
GRU implementation using basic operator, supports multiple layers and bidirectional gru.
.. math::
u_t & = actGate(W_ux xu_{t} + W_uh h_{t-1} + b_u)
r_t & = actGate(W_rx xr_{t} + W_rh h_{t-1} + b_r)
m_t & = actNode(W_cx xm_t + W_ch dot(r_t, h_{t-1}) + b_m)
h_t & = dot(u_t, h_{t-1}) + dot((1-u_t), m_t)
Args:
input (Variable): GRU input tensor,
if batch_first = False, shape should be ( seq_len x batch_size x input_size )
if batch_first = True, shape should be ( batch_size x seq_len x hidden_size )
init_hidden(Variable|None): The initial hidden state of the GRU
This is a tensor with shape ( num_layers x batch_size x hidden_size)
if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
and can be reshaped to tensor with ( num_layers x 2 x batch_size x hidden_size) to use.
If it's None, it will be set to all 0.
hidden_size (int): Hidden size of the GRU
num_layers (int): The total number of layers of the GRU
sequence_length (Variabe|None): A Tensor (shape [batch_size]) stores each real length of each instance,
This tensor will be convert to a mask to mask the padding ids
If it's None means NO padding ids
dropout_prob(float|0.0): Dropout prob, dropout ONLY works after rnn output of each layers,
NOT between time steps
bidirectional (bool|False): If it is bidirectional
batch_first (bool|True): The shape format of the input and output tensors. If true,
the shape format should be :attr:`[batch_size, seq_len, hidden_size]`. If false,
the shape format should be :attr:`[seq_len, batch_size, hidden_size]`. By default
this function accepts input and emits output in batch-major form to be consistent
with most of data format, though a bit less efficient because of extra transposes.
param_attr(ParamAttr|None): The parameter attribute for the learnable
weight matrix. Note:
If it is set to None or one attribute of ParamAttr, gru_unit will
create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|None): The parameter attribute for the bias
of GRU unit.
If it is set to None or one attribute of ParamAttr, gru_unit will
create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
gate_activation (function|None): The activation function for gates (actGate).
Default: 'fluid.layers.sigmoid'
activation (function|None): The activation function for cell (actNode).
Default: 'fluid.layers.tanh'
dtype(string): data type used in this unit
name(string): name used to identify parameters and biases
Returns:
rnn_out(Tensor),last_hidden(Tensor)
- rnn_out is result of GRU hidden, with shape (seq_len x batch_size x hidden_size) \
if is_bidirec set to True, shape will be ( seq_len x batch_sze x hidden_size*2)
- last_hidden is the hidden state of the last step of GRU \
shape is ( num_layers x batch_size x hidden_size ) \
if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size),
can be reshaped to a tensor with shape( num_layers x 2 x batch_size x hidden_size)
Examples:
.. code-block:: python
import paddle.fluid.layers as layers
from paddle.fluid.contrib.layers import basic_gru
batch_size = 20
input_size = 128
hidden_size = 256
num_layers = 2
dropout = 0.5
bidirectional = True
batch_first = False
input = layers.data( name = "input", shape = [-1, batch_size, input_size], dtype='float32')
pre_hidden = layers.data( name = "pre_hidden", shape=[-1, hidden_size], dtype='float32')
sequence_length = layers.data( name="sequence_length", shape=[-1], dtype='int32')
rnn_out, last_hidden = basic_gru( input, pre_hidden, hidden_size, num_layers = num_layers, \
sequence_length = sequence_length, dropout_prob=dropout, bidirectional = bidirectional, \
batch_first = batch_first)
"""
fw_unit_list = []
for i in range(num_layers):
new_name = name + "_layers_" + str(i)
if param_attr is not None and param_attr.name is not None:
layer_param_attr = copy.deepcopy(param_attr)
layer_param_attr.name += "_fw_w_" + str(i)
else:
layer_param_attr = param_attr
if bias_attr is not None and bias_attr.name is not None:
layer_bias_attr = copy.deepcopy(bias_attr)
layer_bias_attr.name += "_fw_b_" + str(i)
else:
layer_bias_attr = bias_attr
fw_unit_list.append(
BasicGRUUnit(new_name, hidden_size, layer_param_attr,
layer_bias_attr, gate_activation, activation, dtype))
if bidirectional:
bw_unit_list = []
for i in range(num_layers):
new_name = name + "_reverse_layers_" + str(i)
if param_attr is not None and param_attr.name is not None:
layer_param_attr = copy.deepcopy(param_attr)
layer_param_attr.name += "_bw_w_" + str(i)
else:
layer_param_attr = param_attr
if bias_attr is not None and bias_attr.name is not None:
layer_bias_attr = copy.deepcopy(bias_attr)
layer_bias_attr.name += "_bw_b_" + str(i)
else:
layer_bias_attr = bias_attr
bw_unit_list.append(
BasicGRUUnit(new_name, hidden_size, layer_param_attr,
layer_bias_attr, gate_activation, activation,
dtype))
if batch_first:
input = layers.transpose(input, [1, 0, 2])
mask = None
if sequence_length:
max_seq_len = layers.shape(input)[0]
mask = layers.sequence_mask(sequence_length,
maxlen=max_seq_len,
dtype='float32')
mask = layers.transpose(mask, [1, 0])
direc_num = 1
if bidirectional:
direc_num = 2
if init_hidden:
init_hidden = layers.reshape(
init_hidden, shape=[num_layers, direc_num, -1, hidden_size])
def get_single_direction_output(rnn_input,
unit_list,
mask=None,
direc_index=0):
rnn = StaticRNN()
with rnn.step():
step_input = rnn.step_input(rnn_input)
if mask:
step_mask = rnn.step_input(mask)
for i in range(num_layers):
if init_hidden:
pre_hidden = rnn.memory(init=init_hidden[i, direc_index])
else:
pre_hidden = rnn.memory(batch_ref=rnn_input,
shape=[-1, hidden_size],
ref_batch_dim_idx=1)
new_hidden = unit_list[i](step_input, pre_hidden)
if mask:
new_hidden = layers.elementwise_mul(
new_hidden, step_mask,
axis=0) - layers.elementwise_mul(pre_hidden,
(step_mask - 1),
axis=0)
rnn.update_memory(pre_hidden, new_hidden)
rnn.step_output(new_hidden)
step_input = new_hidden
if dropout_prob != None and dropout_prob > 0.0:
step_input = layers.dropout(
step_input,
dropout_prob=dropout_prob,
)
rnn.step_output(step_input)
rnn_out = rnn()
last_hidden_array = []
rnn_output = rnn_out[-1]
for i in range(num_layers):
last_hidden = rnn_out[i]
last_hidden = last_hidden[-1]
last_hidden_array.append(last_hidden)
last_hidden_output = layers.concat(last_hidden_array, axis=0)
last_hidden_output = layers.reshape(
last_hidden_output, shape=[num_layers, -1, hidden_size])
return rnn_output, last_hidden_output
# seq_len, batch_size, hidden_size
fw_rnn_out, fw_last_hidden = get_single_direction_output(input,
fw_unit_list,
mask,
direc_index=0)
if bidirectional:
bw_input = layers.reverse(input, axis=[0])
bw_mask = None
if mask:
bw_mask = layers.reverse(mask, axis=[0])
bw_rnn_out, bw_last_hidden = get_single_direction_output(bw_input,
bw_unit_list,
bw_mask,
direc_index=1)
bw_rnn_out = layers.reverse(bw_rnn_out, axis=[0])
rnn_out = layers.concat([fw_rnn_out, bw_rnn_out], axis=2)
last_hidden = layers.concat([fw_last_hidden, bw_last_hidden], axis=1)
last_hidden = layers.reshape(
last_hidden, shape=[num_layers * direc_num, -1, hidden_size])
if batch_first:
rnn_out = layers.transpose(rnn_out, [1, 0, 2])
return rnn_out, last_hidden
else:
rnn_out = fw_rnn_out
last_hidden = fw_last_hidden
if batch_first:
rnn_out = layers.transpose(rnn_out, [1, 0, 2])
return rnn_out, last_hidden
def forward(self,
inputs,
initial_states=None,
sequence_length=None,
**kwargs):
if F.in_dygraph_mode():
class OutputArray(object):
def __init__(self, x):
self.array = [x]
def append(self, x):
self.array.append(x)
def _maybe_copy(state, new_state, step_mask):
# TODO: use where_op
new_state = L.elementwise_mul(new_state, step_mask, axis=0) - \
L.elementwise_mul(state, (step_mask - 1), axis=0)
return new_state
#logging.info("inputs shape: {}".format(inputs.shape))
flat_inputs = U.flatten(inputs)
#logging.info("flat inputs len: {}".format(len(flat_inputs)))
#logging.info("flat inputs[0] shape: {}".format(flat_inputs[0].shape))
batch_size, time_steps = (
flat_inputs[0].shape[self.batch_index],
flat_inputs[0].shape[self.time_step_index])
#logging.info("batch_size: {}".format(batch_size))
#logging.info("time_steps: {}".format(time_steps))
if initial_states is None:
initial_states = self.cell.get_initial_states(
batch_ref=inputs, batch_dim_idx=self.batch_index)
if not self.time_major:
# 如果第一维不是时间步 则第一维和第二维交换
# 第一维为时间步
inputs = U.map_structure(
lambda x: L.transpose(x, [1, 0] + list(
range(2, len(x.shape)))), inputs)
if sequence_length is not None:
mask = L.sequence_mask(
sequence_length,
maxlen=time_steps,
dtype=U.flatten(initial_states)[0].dtype)
# 同样 第一维为时间步
mask = L.transpose(mask, [1, 0])
if self.is_reverse:
# 如果反向
# 则第一维反向
inputs = U.map_structure(lambda x: L.reverse(x, axis=[0]), inputs)
mask = L.reverse(mask, axis=[0]) if sequence_length is not None else None
states = initial_states
outputs = []
# 遍历时间步
for i in range(time_steps):
# 取该时间步的输入
step_inputs = U.map_structure(lambda x: x[i], inputs)
# 输入当前输入和状态
# 得到输出和新状态
step_outputs, new_states = self.cell(step_inputs, states, **kwargs)
if sequence_length is not None:
# 如果有mask 则被mask的地方 用原state的数
# _maybe_copy: 未mask的部分用new_states, 被mask的部分用states
new_states = U.map_structure(
partial(_maybe_copy, step_mask=mask[i]),
states,
new_states)
states = new_states
#logging.info("step_output shape: {}".format(step_outputs.shape))
if i == 0:
# 初始时，各输出
outputs = U.map_structure(lambda x: OutputArray(x), step_outputs)
else:
# 各输出加入对应list中
U.map_structure(lambda x, x_array: x_array.append(x), step_outputs, outputs)
# 最后按时间步的维度堆叠
final_outputs = U.map_structure(
lambda x: L.stack(x.array, axis=self.time_step_index),
outputs)
#logging.info("final_outputs shape: {}".format(final_outputs.shape))
if self.is_reverse:
# 如果是反向 则最后结果也反向一下
final_outputs = U.map_structure(
lambda x: L.reverse(x, axis=self.time_step_index),
final_outputs)
final_states = new_states
else:
final_outputs, final_states = L.rnn(
self.cell,
inputs,
initial_states=initial_states,
sequence_length=sequence_length,
time_major=self.time_major,
is_reverse=self.is_reverse,
**kwargs)
return final_outputs, final_states
def transform(self, img, do_flip):
if do_flip:
if isinstance(img, PTensor):
return layers.reverse(img, 2)
return np.fliplr(img).copy()
return img
def conditional_gru(input,
encode_hidden,
init_hidden,
encode_hidden_size,
hidden_size,
num_layers=1,
sequence_length=None,
dropout_prob=0.0,
bidirectional=False,
batch_first=True,
param_attr=None,
bias_attr=None,
gate_activation=None,
activation=None,
dtype="float32",
name="conditional_gru"):
"""
定义一个新的GRU类型，多了参数Cu,Cr,C。GRU的新公式:
.. math::
u_t & = actGate(W_ux xu_{t} + W_uh h_{t-1} + C_u h_i + b_u)
r_t & = actGate(W_rx xr_{t} + W_rh h_{t-1} + C_r h_i + b_r)
m_t & = actNode(W_cx xm_t + W_ch dot(r_t, h_{t-1}) + C_u h_i + C h_i + b_m)
h_t & = dot(u_t, h_{t-1}) + dot((1-u_t), m_t)
其他定义与GRU相同
Args:
input (Variable): GRU input tensor,
if batch_first = False, shape should be ( seq_len x batch_size x input_size )
if batch_first = True, shape should be ( batch_size x seq_len x hidden_size )
encode_hidden: The hidden state from the encoder of the GRU. If bidirectional is True, the encode_hidden is assert
to contain two parts, former half part is for forward direction, and later half part is for backward
direction.
encode_hidden_size: The size of encode_hidden. If bidirectional is True, the encode_hidden_size includes the
former half part and the later half part, i.e., the actual size of encode_hidden is
encode_hidden_size / 2
init_hidden(Variable|None): The initial hidden state of the GRU
This is a tensor with shape ( num_layers x batch_size x hidden_size)
if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
and can be reshaped to tensor with ( num_layers x 2 x batch_size x hidden_size) to use.
If it's None, it will be set to all 0.
hidden_size (int): Hidden size of the GRU
num_layers (int): The total number of layers of the GRU
sequence_length (Variabe|None): A Tensor (shape [batch_size]) stores each real length of each instance,
This tensor will be convert to a mask to mask the padding ids
If it's None means NO padding ids
dropout_prob(float|0.0): Dropout prob, dropout ONLY works after rnn output of each layers,
NOT between time steps
bidirectional (bool|False): If it is bidirectional
batch_first (bool|True): The shape format of the input and output tensors. If true,
the shape format should be :attr:`[batch_size, seq_len, hidden_size]`. If false,
the shape format should be :attr:`[seq_len, batch_size, hidden_size]`. By default
this function accepts input and emits output in batch-major form to be consistent
with most of data format, though a bit less efficient because of extra transposes.
param_attr(ParamAttr|None): The parameter attribute for the learnable
weight matrix. Note:
If it is set to None or one attribute of ParamAttr, gru_unit will
create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|None): The parameter attribute for the bias
of GRU unit.
If it is set to None or one attribute of ParamAttr, gru_unit will
create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
gate_activation (function|None): The activation function for gates (actGate).
Default: 'fluid.layers.sigmoid'
activation (function|None): The activation function for cell (actNode).
Default: 'fluid.layers.tanh'
dtype(string): data type used in this unit
name(string): name used to identify parameters and biases
Returns:
rnn_out(Tensor),last_hidden(Tensor)
- rnn_out is result of GRU hidden, with shape (seq_len x batch_size x hidden_size) \
if is_bidirec set to True, shape will be ( seq_len x batch_sze x hidden_size*2)
- last_hidden is the hidden state of the last step of GRU \
shape is ( num_layers x batch_size x hidden_size ) \
if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size),
can be reshaped to a tensor with shape( num_layers x 2 x batch_size x hidden_size)
- all_hidden is all the hidden states of the input, including the last_hidden and medium hidden states. \
shape is (num_layers x seq_len x batch_size x hidden_size). if is_bidirec set to True, shape will be
(2 x num_layers x seq_len x batch_size x hidden_size)
"""
if bidirectional:
encode_hidden, bw_encode_hidden = layers.split(encode_hidden, num_or_sections=2, dim=-1)
encode_hidden_size = int(encode_hidden_size / 2)
fw_unit_list = []
for i in range(num_layers):
new_name = name + '_layers_' + str(i)
if param_attr is not None and param_attr.name is not None:
layer_param_attr = copy.deepcopy(param_attr)
layer_param_attr.name += '_fw_w_' + str(i)
else:
layer_param_attr = param_attr
if bias_attr is not None and bias_attr.name is not None:
layer_bias_attr = copy.deepcopy(bias_attr)
layer_bias_attr.name += '_fw_b_' + str(i)
else:
layer_bias_attr = bias_attr
fw_unit_list.append(
ConditionalGRUUnit(new_name, encode_hidden_size, hidden_size, layer_param_attr,
layer_bias_attr, gate_activation, activation, dtype)
)
if bidirectional:
bw_unit_list = []
for i in range(num_layers):
new_name = name + '_reverse_layers_' + str(i)
if param_attr is not None and param_attr.name is not None:
layer_param_attr = copy.deepcopy(param_attr)
layer_param_attr.name += '_bw_w_' + str(i)
else:
layer_param_attr = param_attr
if bias_attr is not None and bias_attr.name is not None:
layer_bias_attr = copy.deepcopy(bias_attr)
layer_bias_attr.name += '_bw_b_' + str(i)
else:
layer_bias_attr = bias_attr
bw_unit_list.append(
ConditionalGRUUnit(new_name, encode_hidden_size, hidden_size, layer_param_attr,
layer_bias_attr, gate_activation, activation, dtype)
)
if batch_first:
input = layers.transpose(input, [1, 0, 2])
mask = None
if sequence_length:
max_seq_len = layers.shape(input)[0]
mask = layers.sequence_mask(
sequence_length, maxlen=max_seq_len, dtype='float32'
)
mask = layers.transpose(mask, [1, 0])
direc_num = 1
if bidirectional:
direc_num = 2
if init_hidden:
init_hidden = layers.reshape(
init_hidden, shape=[num_layers, direc_num, -1, hidden_size]
)
def get_single_direction_output(rnn_input,
encode_hidden,
unit_list,
mask=None,
direc_index=0):
rnn = StaticRNN()
#print(rnn_input.shape)
with rnn.step():
step_input = rnn.step_input(rnn_input)
if mask:
step_mask = rnn.step_input(mask)
for i in range(num_layers):
if init_hidden:
pre_hidden = rnn.memory(init=init_hidden[i, direc_index])
else:
pre_hidden = rnn.memory(batch_ref=rnn_input,
shape=[-1, hidden_size],
ref_batch_dim_idx=1)
encode_h = encode_hidden[i]
pre_encode_hidden = layers.concat([pre_hidden, encode_h], axis=1)
new_hidden = unit_list[i](step_input, pre_encode_hidden)
if mask:
new_hidden = layers.elementwise_mul(
new_hidden, step_mask, axis=0) - layers.elementwise_mul(
pre_hidden, (step_mask - 1), axis=0)
rnn.update_memory(pre_hidden, new_hidden)
rnn.step_output(new_hidden)
step_input = new_hidden
if dropout_prob is not None and dropout_prob > 0.0:
step_input = layers.dropout(step_input, dropout_prob=dropout_prob, )
rnn.step_output(step_input)
rnn_out = rnn()
last_hidden_array = []
all_hidden_array = [] # 增加这个来得到所有隐含状态
rnn_output = rnn_out[-1]
for i in range(num_layers):
last_hidden = rnn_out[i]
all_hidden_array.append(last_hidden)
last_hidden = last_hidden[-1]
last_hidden_array.append(last_hidden)
all_hidden_array = layers.concat(all_hidden_array, axis=0)
all_hidden_array = layers.reshape(all_hidden_array, shape=[num_layers, input.shape[0], -1, hidden_size])
last_hidden_output = layers.concat(last_hidden_array, axis=0)
last_hidden_output = layers.reshape(last_hidden_output, shape=[num_layers, -1, hidden_size])
return rnn_output, last_hidden_output, all_hidden_array
fw_rnn_out, fw_last_hidden, fw_all_hidden = get_single_direction_output(
input, encode_hidden, fw_unit_list, mask, direc_index=0)
if bidirectional:
bw_input = layers.reverse(input, axis=[0])
bw_mask = None
if mask:
bw_mask = layers.reverse(mask, axis=[0])
bw_rnn_out, bw_last_hidden, bw_all_hidden = get_single_direction_output(
bw_input, bw_encode_hidden, bw_unit_list, bw_mask, direc_index=1)
bw_rnn_out = layers.reverse(bw_rnn_out, axis=[0])
rnn_out = layers.concat([fw_rnn_out, bw_rnn_out], axis=2)
last_hidden = layers.concat([fw_last_hidden, bw_last_hidden], axis=1)
all_hidden = layers.concat([fw_all_hidden, bw_all_hidden], axis=0)
last_hidden = layers.reshape(
last_hidden, shape=[num_layers * direc_num, -1, hidden_size])
if batch_first:
rnn_out = layers.transpose(rnn_out, [1, 0, 2])
return rnn_out, last_hidden, all_hidden
else:
rnn_out = fw_rnn_out
last_hidden = fw_last_hidden
all_hidden = fw_all_hidden
if batch_first:
rnn_out = layers.transpose(rnn_out, [1, 0, 2])
return rnn_out, last_hidden, all_hidden
def __call__(self, image: PTensor):
if isinstance(image, PTensor):
return self.crop_to_output(layers.reverse(image, 2))
else:
return self.crop_to_output(np.flipud(image))