def F(inputs, d, activation=tf.nn.relu, kernel_initializer=None, scope=None, use_bias=True, input_keep_prob=1.0, wd=0.0, is_train=None):
out = dropout(inputs, input_keep_prob, is_train)
with tf.variable_scope(scope or "projection"):
out = tf.layers.dense(out, d, activation=activation, use_bias=use_bias, kernel_initializer=kernel_initializer)
if wd:
add_wd(wd)
return out
def linear(args,
output_size,
bias,
bias_start=0.0,
scope=None,
squeeze=False,
wd=0.0,
input_keep_prob=1.0,
is_train=None,
kernel_initializer=None):
if args is None or (nest.is_sequence(args) and not args):
raise ValueError("`args` must be specified")
if not nest.is_sequence(args):
args = [args]
flat_args = [flatten(arg, 1)
for arg in args] # flat_args[0] : [N*JX*JQ, d]
if input_keep_prob < 1.0:
assert is_train is not None
flat_args = [
tf.cond(is_train, lambda: tf.nn.dropout(arg, input_keep_prob),
lambda: arg) for arg in flat_args
]
with tf.variable_scope(scope or 'linear'):
flat_out = _linear(flat_args,
output_size,
bias,
kernel_initializer=kernel_initializer)
out = reconstruct(flat_out, args[0], 1)
if squeeze:
out = tf.squeeze(out, [len(args[0].get_shape().as_list()) - 1])
if wd:
add_wd(wd)
return out
def conv2d(in_,
filter_size,
height,
padding,
dilation_rate=None,
wd=0.0,
is_train=None,
keep_prob=1.0,
scope=None,
nonlinear=True):
with tf.variable_scope(scope or "conv2d"):
num_channels = in_.get_shape().as_list()[-1]
N = tf.shape(in_)[0]
in_ = tf.expand_dims(in_, 1)
filter_ = tf.get_variable("filter_",
shape=[1, height, num_channels, filter_size],
dtype='float')
bias = tf.get_variable("bias", shape=[filter_size], dtype='float')
strides = [1, 1]
dilation_rate = [1, dilation_rate or 1]
if is_train is not None and keep_prob < 1.0:
in_ = dropout(in_, keep_prob, is_train)
out = tf.nn.convolution(in_,
filter_,
padding,
strides=strides,
dilation_rate=dilation_rate) + bias
if nonlinear:
out = tf.nn.tanh(out)
out = tf.reshape(out, [N, -1, filter_size])
if wd:
add_wd(wd)
return out
def conv1d(in_, filter_size, height, padding, wd=0.0, is_train=None, keep_prob=1.0, scope=None):
with tf.variable_scope(scope or "conv1d"):
num_channels = in_.get_shape()[-1]
filter_ = tf.get_variable("filter", shape=[1, height, num_channels, filter_size], dtype='float')
bias = tf.get_variable("bias", shape=[filter_size], dtype='float')
strides = [1, 1, 1, 1]
if is_train is not None and keep_prob < 1.0:
in_ = dropout(in_, keep_prob, is_train)
xxc = tf.nn.conv2d(in_, filter_, strides, padding) + bias # [N*M, JX, W/filter_stride, d]
out = tf.reduce_max(tf.nn.relu(xxc), 2) # [-1, JX, d]
if wd:
add_wd(wd)
return out
def rnn(rnn_type, inputs, length, hidden_size, num_layers=1, state_keep_prob=1.0,
dropout_keep_prob=None, concat=True, initial_state=None,
kernel_initializer=tf.random_normal_initializer(stddev=0.1), wd=0.0, is_train=False, scope=None):
with tf.variable_scope(scope or 'rnn'):
if not rnn_type.startswith('bi'):
cell = get_cell(rnn_type, hidden_size, num_layers, dropout_keep_prob, kernel_initializer=kernel_initializer, is_train=is_train)
outputs, states = tf.nn.dynamic_rnn(cell, inputs, sequence_length=length, dtype=tf.float32, initial_state=initial_state)
if rnn_type.endswith('lstm'):
h = states[0][1]
# h = [state.h for state in states]
state = h
else:
cell_fw = get_cell(rnn_type, hidden_size, num_layers, dropout_keep_prob, state_keep_prob, kernel_initializer=kernel_initializer, is_train=is_train)
cell_bw = get_cell(rnn_type, hidden_size, num_layers, dropout_keep_prob, state_keep_prob, kernel_initializer=kernel_initializer, is_train=is_train)
if initial_state is not None:
outputs, states = tf.nn.bidirectional_dynamic_rnn(
cell_bw, cell_fw, inputs, sequence_length=length, dtype=tf.float32,
initial_state_fw=initial_state[:, :hidden_size], initial_state_bw=initial_state[:, hidden_size:]
)
else:
outputs, states = tf.nn.bidirectional_dynamic_rnn(
cell_bw, cell_fw, inputs, sequence_length=length, dtype=tf.float32
)
state_fw, state_bw = states
if rnn_type.endswith('lstm'):
h_fw = state_fw[0][1]
h_bw = state_bw[0][1]
# h_fw = [state_fw.h for state_fw in states_fw]
# h_bw = [state_bw.h for state_bw in states_bw]
state_fw, state_bw = h_fw, h_bw
if concat:
outputs = tf.concat(outputs, 2)
state = tf.concat([state_fw, state_bw], 1)
else:
outputs = outputs[0] + outputs[1]
state = state_fw + state_bw
if wd:
add_wd(wd)
return outputs, state
def cudnn_rnn(rnn_type, inputs, length, hidden_size, num_layers=1,
dropout_keep_prob=1.0, concat=True, initial_state=None,
kernel_initializer=tf.random_normal_initializer(stddev=0.1), wd=0.0, is_train=False, scope=None):
with tf.variable_scope(scope or 'cudnn_rnn'):
direction = "bidirectional" if 'bi' in rnn_type else "unidirectional"
input_size = inputs.get_shape().as_list()[-1]
if rnn_type.endswith('gru'):
rnn = CudnnGRU(num_layers=num_layers, num_units=hidden_size,
input_mode='linear_input', direction=direction,
dropout=1-dropout_keep_prob, name='rnn')
elif rnn_type.endswith('lstm'):
rnn = CudnnLSTM(num_layers=num_layers, num_units=hidden_size,
input_mode='linear_input', direction=direction,
dropout=1-dropout_keep_prob, name='rnn')
else:
raise NotImplementedError("{} is not supported.".format(rnn_type))
inputs = dropout(inputs, dropout_keep_prob, is_train)
outputs, _ = rnn(tf.transpose(inputs, [1, 0, 2]))
outputs = tf.transpose(outputs, [1, 0, 2]) # [N, JX, 2*d]
output_h = None
if wd:
add_wd(wd)
return outputs, output_h
