def __call__(self, inputs, inputs_seq_len,
keep_prob_input, keep_prob_hidden, keep_prob_output):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size`[B, T, input_size]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob_input (placeholder, float): A probability to keep nodes
in the input-hidden connection
keep_prob_hidden (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
keep_prob_output (placeholder, float): A probability to keep nodes
in the hidden-output connection
Returns:
logits: A tensor of size `[T, B, num_classes]`
final_state: A final hidden state of the encoder
"""
# inputs: 3D tensor `[batch_size, max_time, input_size * splice]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
# Reshape to 4D tensor
# `[batch_size * max_time, input_size / 3, splice, 3(+Δ,ΔΔ)]`
inputs = tf.reshape(
inputs,
shape=[batch_size * max_time, int(self.input_size / 3), self.splice, 3])
with tf.variable_scope('VGG1'):
inputs = conv_layer(inputs,
filter_shape=[3, 3, 3, 96],
parameter_init=self.parameter_init,
relu=True,
name='conv1')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 96, 96],
parameter_init=self.parameter_init,
relu=True,
name='conv2')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 96, 96],
parameter_init=self.parameter_init,
relu=True,
name='conv3')
inputs = max_pool(inputs, name='max_pool')
# TODO(hirofumi): try batch normalization
with tf.variable_scope('VGG2'):
inputs = conv_layer(inputs,
filter_shape=[3, 3, 96, 192],
parameter_init=self.parameter_init,
relu=True,
name='conv1')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 192, 192],
parameter_init=self.parameter_init,
relu=True,
name='conv2')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 192, 192],
parameter_init=self.parameter_init,
relu=True,
name='conv3')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 192, 192],
parameter_init=self.parameter_init,
relu=True,
name='conv4')
inputs = max_pool(inputs, name='max_pool')
# TODO(hirofumi): try batch normalization
with tf.variable_scope('VGG3'):
inputs = conv_layer(inputs,
filter_shape=[3, 3, 192, 384],
parameter_init=self.parameter_init,
relu=True,
name='conv1')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 384, 384],
parameter_init=self.parameter_init,
relu=True,
name='conv2')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 384, 384],
parameter_init=self.parameter_init,
relu=True,
name='conv3')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 384, 384],
parameter_init=self.parameter_init,
relu=True,
name='conv4')
inputs = max_pool(inputs, name='max_pool')
# TODO(hirofumi): try batch normalization
# Reshape to 2D tensor `[batch_size * max_time, new_h * new_w * 384]`
new_h = math.ceil(self.input_size / (3 * 2**3)
) # expected to be 5 or 6
new_w = math.ceil(self.splice / (2**3)) # expected to be 2
inputs = tf.reshape(
inputs, shape=[batch_size * max_time, new_h * new_w * 384])
with tf.variable_scope('fc1') as scope:
inputs = tf.contrib.layers.fully_connected(
inputs=inputs,
num_outputs=1024,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
with tf.variable_scope('fc2') as scope:
inputs = tf.contrib.layers.fully_connected(
inputs=inputs,
num_outputs=1024,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
with tf.variable_scope('fc3') as scope:
logits_2d = tf.contrib.layers.fully_connected(
inputs=inputs,
num_outputs=self.num_classes,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
# if self.bottleneck_dim is not None and self.bottleneck_dim != 0:
# with tf.variable_scope('bottleneck') as scope:
# outputs = tf.contrib.layers.fully_connected(
# outputs, self.bottleneck_dim,
# activation_fn=tf.nn.relu,
# weights_initializer=tf.truncated_normal_initializer(stddev=0.1),
# biases_initializer=tf.zeros_initializer(),
# scope=scope)
#
# # Dropout for the hidden-output connections
# outputs = tf.nn.dropout(
# outputs, keep_prob_output, name='dropout_output_bottle')
# Reshape back to 3D tensor `[batch_size, max_time, num_classes]`
logits = tf.reshape(
logits_2d, shape=[batch_size, max_time, self.num_classes])
# Convert to time-major: `[max_time, batch_size, num_classes]'
logits = tf.transpose(logits, (1, 0, 2))
return logits, None
def __call__(self, inputs, inputs_seq_len, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, T, input_size (num_channels * splice * num_stack * 3)]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size `[T, B, output_dim]`
otherwise, `[B, T, output_dim]`
final_state: None
"""
# inputs: 3D tensor `[B, T, input_dim]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
# For debug
# print(input_dim)
# print(self.num_channels)
# print(self.splice)
# print(self.num_stack)
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B * T, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(
inputs,
shape=[batch_size * max_time, self.num_channels, self.splice * self.num_stack, 3])
# NOTE: filter_size: `[H, W, C_in, C_out]`
with tf.variable_scope('CNN1'):
inputs = conv_layer(inputs,
filter_size=[9, 9, 3, 64],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[3, 1],
stride=[3, 1],
name='max_pool')
# TODO: try dropout
with tf.variable_scope('CNN2'):
inputs = conv_layer(inputs,
filter_size=[3, 4, 64, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
# TODO: try dropout
# Reshape to 2D tensor `[B * T, new_h * new_w * C_out]`
outputs = tf.reshape(
inputs, shape=[batch_size * max_time, np.prod(inputs.shape.as_list()[-3:])])
for i in range(1, 3, 1):
with tf.variable_scope('fc%d' % (i)) as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=768,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
outputs = tf.nn.dropout(outputs, keep_prob)
# Reshape back to 3D tensor `[B, T, 768]`
outputs = tf.reshape(
outputs, shape=[batch_size, max_time, 768])
if self.time_major:
# Convert to time-major: `[T, B, num_classes]'
outputs = tf.transpose(outputs, [1, 0, 2])
return outputs, None
def __call__(self, inputs, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, input_size (num_channels * splice * num_stack * 3)]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size `[T, B, output_dim]`
otherwise, `[B, output_dim]`
"""
# inputs: 2D tensor `[B, input_dim]`
batch_size = tf.shape(inputs)[0]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
# for debug
# print(input_dim) # 1200
# print(self.num_channels) # 40
# print(self.splice) # 5
# print(self.num_stack) # 2
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(inputs,
shape=[
batch_size, self.num_channels,
self.splice * self.num_stack, 3
])
# NOTE: filter_size: `[H, W, C_in, C_out]`
with tf.variable_scope('CNN1'):
inputs = conv_layer(inputs,
filter_size=[9, 9, 3, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[3, 1],
stride=[3, 1],
name='max_pool')
with tf.variable_scope('CNN2'):
inputs = conv_layer(inputs,
filter_size=[3, 4, 128, 256],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
# Reshape to 2D tensor `[B, new_h * new_w * C_out]`
outputs = tf.reshape(
inputs, shape=[batch_size,
np.prod(inputs.shape.as_list()[-3:])])
for i in range(1, 5, 1):
with tf.variable_scope('fc%d' % (i)) as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=2048,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
return outputs
def __call__(self, inputs, inputs_seq_len, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, T, input_size (num_channels * (splice * num_stack) * 3)]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size
`[T, B, num_units (num_proj)]`
otherwise, `[B, T, num_units (num_proj)]`
final_state: A final hidden state of the encoder
"""
# inputs: 3D tensor `[B, T, input_dim]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
# For debug
# print(input_dim)
# print(self.num_channels)
# print(self.splice)
# print(self.num_stack)
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B * T, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(
inputs,
shape=[batch_size * max_time, self.num_channels, self.splice * self.num_stack, 3])
# NOTE: filter_size: `[H, W, C_in, C_out]`
with tf.variable_scope('CNN1'):
inputs = conv_layer(inputs,
filter_size=[11, 21, 3, 32],
stride=[3, 2],
parameter_init=self.parameter_init,
activation='relu')
# inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
with tf.variable_scope('CNN2'):
inputs = conv_layer(inputs,
filter_size=[11, 11, 32, 32],
stride=[1, 2],
parameter_init=self.parameter_init,
activation='relu')
# inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
with tf.variable_scope('CNN3'):
inputs = conv_layer(inputs,
filter_size=[3, 3, 32, 96],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
# inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
# Reshape to 3D tensor `[B, T, new_h * new_w * C_out]`
inputs = tf.reshape(
inputs, shape=[batch_size, max_time, np.prod(inputs.shape.as_list()[-3:])])
initializer = tf.random_uniform_initializer(
minval=-self.parameter_init, maxval=self.parameter_init)
if self.lstm_impl == 'BasicLSTMCell':
outputs, final_state = basiclstmcell(
self.num_units, self.num_layers,
inputs, inputs_seq_len, keep_prob, initializer,
self.time_major)
elif self.lstm_impl == 'LSTMCell':
outputs, final_state = lstmcell(
self.num_units, self.num_proj, self.num_layers,
self.use_peephole, self.clip_activation,
inputs, inputs_seq_len, keep_prob, initializer,
self.time_major)
elif self.lstm_impl == 'LSTMBlockCell':
outputs, final_state = lstmblockcell(
self.num_units, self.num_layers,
self.use_peephole, self.clip_activation,
inputs, inputs_seq_len, keep_prob, initializer,
self.time_major)
elif self.lstm_impl == 'LSTMBlockFusedCell':
outputs, final_state = lstmblockfusedcell(
self.num_units, self.num_layers,
inputs, inputs_seq_len, keep_prob, initializer,
self.time_major)
elif self.lstm_impl == 'CudnnLSTM':
outputs, final_state = cudnnlstm(
self.num_units, self.num_layers,
inputs, inputs_seq_len, keep_prob, initializer,
self.time_major)
else:
raise IndexError(
'lstm_impl is "BasicLSTMCell" or "LSTMCell" or ' +
'"LSTMBlockCell" or "LSTMBlockFusedCell" or ' +
'"CudnnLSTM".')
# Reshape to 2D tensor `[B * T (T * B), output_dim]`
output_dim = outputs.shape.as_list()[-1]
outputs = tf.reshape(
outputs, shape=[batch_size * max_time, output_dim])
with tf.variable_scope('fc1') as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=896,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
outputs = tf.nn.dropout(outputs, keep_prob)
with tf.variable_scope('fc2') as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=74,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
output_dim = outputs.shape.as_list()[-1]
if self.time_major:
# Reshape back to 3D tensor `[T, B, 74]`
outputs = tf.reshape(
outputs, shape=[max_time, batch_size, output_dim])
else:
# Reshape back to 3D tensor `[B, T, 74]`
outputs = tf.reshape(
outputs, shape=[batch_size, max_time, output_dim])
return outputs, final_state
def __call__(self, inputs, inputs_seq_len, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, T, input_size (num_channels * (splice * num_stack) * 3)]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size `[T, B, output_dim]`
otherwise, `[B, T, output_dim]`
final_state: None
"""
# inputs: 3D tensor `[B, T, input_dim]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
# For debug
# print(input_dim)
# print(self.num_channels)
# print(self.splice)
# print(self.num_stack)
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B * T, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(
inputs,
shape=[batch_size * max_time, self.num_channels, self.splice * self.num_stack, 3])
# NOTE: filter_size: `[H, W, C_in, C_out]`
with tf.variable_scope('VGG1'):
for i_layer in range(1, 4, 1):
input_channels = inputs.shape.as_list()[-1]
inputs = conv_layer(inputs,
filter_size=[3, 3, input_channels, 96],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu',
name='conv1')
inputs = batch_normalization(inputs, is_training=is_training)
if i_layer == 3:
inputs = max_pool(inputs,
pooling_size=[2, 2],
stride=[2, 2],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
with tf.variable_scope('VGG2'):
for i_layer in range(1, 5, 1):
input_channels = inputs.shape.as_list()[-1]
inputs = conv_layer(inputs,
filter_size=[3, 3, input_channels, 192],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu',
name='conv%d' % i_layer)
inputs = batch_normalization(inputs, is_training=is_training)
if i_layer == 4:
inputs = max_pool(inputs,
pooling_size=[2, 2],
stride=[2, 2],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
with tf.variable_scope('VGG3'):
for i_layer in range(1, 5, 1):
input_channels = inputs.shape.as_list()[-1]
inputs = conv_layer(inputs,
filter_size=[3, 3, input_channels, 384],
parameter_init=self.parameter_init,
activation='relu',
name='conv%d' % i_layer)
inputs = batch_normalization(inputs, is_training=is_training)
if i_layer == 4:
inputs = max_pool(inputs,
pooling_size=[2, 2],
stride=[2, 2],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
# Reshape to 2D tensor `[B * T, new_h * new_w * C_out]`
outputs = tf.reshape(
inputs, shape=[batch_size * max_time, np.prod(inputs.shape.as_list()[-3:])])
for i_layer in range(1, 3, 1):
with tf.variable_scope('fc%d' % i_layer) as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=1024,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
if i_layer == 1:
outputs = tf.nn.dropout(outputs, keep_prob)
# Reshape back to 3D tensor `[B, T, 1024]`
output_dim = outputs.shape.as_list()[-1]
outputs = tf.reshape(
outputs, shape=[batch_size, max_time, output_dim])
if self.time_major:
# Convert to time-major: `[T, B, num_classes]'
outputs = tf.transpose(outputs, [1, 0, 2])
return outputs, None
def __call__(self, inputs, inputs_seq_len, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, T, input_size (num_channels * splice * num_stack * 3)]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size `[T, B, output_dim]`
otherwise, `[B, T, output_dim]`
final_state: None
"""
# inputs: 3D tensor `[B, T, input_dim]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
# For debug
# print(input_dim)
# print(self.num_channels)
# print(self.splice)
# print(self.num_stack)
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B * T, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(
inputs,
shape=[batch_size * max_time, self.num_channels, self.splice * self.num_stack, 3])
# Choose the activation function
activation = 'relu'
# activation = 'prelu'
# activation = 'maxout'
# TODO: add prelu and maxout layers
# NOTE: filter_size: `[H, W, C_in, C_out]`
# 1-4th layers
for i_layer in range(1, 5, 1):
with tf.variable_scope('CNN%d' % i_layer):
input_channels = inputs.shape.as_list()[-1]
inputs = conv_layer(inputs,
filter_size=[3, 5, input_channels, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation=activation,
name='conv')
# inputs = batch_normalization(inputs, is_training=is_training)
if i_layer == 1:
inputs = max_pool(inputs,
pooling_size=[3, 1],
stride=[3, 1],
name='pool')
inputs = tf.nn.dropout(inputs, keep_prob)
# 5-10th layers
for i_layer in range(5, 11, 1):
with tf.variable_scope('CNN%d' % i_layer):
input_channels = inputs.shape.as_list()[-1]
inputs = conv_layer(inputs,
filter_size=[3, 5, input_channels, 256],
stride=[1, 1],
parameter_init=self.parameter_init,
activation=activation,
name='conv')
# inputs = batch_normalization(inputs, is_training=is_training)
# NOTE: No poling
inputs = tf.nn.dropout(inputs, keep_prob)
# Reshape to 2D tensor `[B * T, new_h * new_w * C_out]`
outputs = tf.reshape(
inputs, shape=[batch_size * max_time, np.prod(inputs.shape.as_list()[-3:])])
# 11-14th layers
for i_layer in range(1, 4, 1):
with tf.variable_scope('fc%d' % i_layer) as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=1024,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
if i_layer != 3:
outputs = tf.nn.dropout(outputs, keep_prob)
# Reshape back to 3D tensor `[B, T, 1024]`
logits = tf.reshape(
outputs, shape=[batch_size, max_time, 1024])
if self.time_major:
# Convert to time-major: `[T, B, 1024]'
logits = tf.transpose(logits, [1, 0, 2])
return logits, None
def __call__(self, inputs, inputs_seq_len,
keep_prob_input, keep_prob_hidden, keep_prob_output):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size`[B, T, input_size]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob_input (placeholder, float): A probability to keep nodes
in the input-hidden connection
keep_prob_hidden (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
keep_prob_output (placeholder, float): A probability to keep nodes
in the hidden-output connection
Returns:
logits: A tensor of size `[T, B, num_classes]`
final_state: A final hidden state of the encoder
"""
# inputs: `[B, T, input_size * splice]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
# Reshape to 4D tensor `[B * T, input_size / 3, splice, 3(+Δ, ΔΔ)]`
inputs = tf.reshape(
inputs,
shape=[batch_size * max_time, int(self.input_size / 3), self.splice, 3])
with tf.variable_scope('VGG1'):
inputs = conv_layer(inputs,
filter_shape=[3, 3, 3, 64],
parameter_init=self.parameter_init,
relu=True,
name='conv1')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 64, 64],
parameter_init=self.parameter_init,
relu=True,
name='conv2')
inputs = max_pool(inputs, name='max_pool')
# TODO(hirofumi): try batch normalization
with tf.variable_scope('VGG2'):
inputs = conv_layer(inputs,
filter_shape=[3, 3, 64, 128],
parameter_init=self.parameter_init,
relu=True,
name='conv1')
inputs = conv_layer(inputs,
filter_shape=[3, 3, 128, 128],
parameter_init=self.parameter_init,
relu=True,
name='conv2')
inputs = max_pool(inputs, name='max_pool')
# TODO(hirofumi): try batch normalization
# Reshape to 2D tensor `[B * T, new_h * new_w * 128]`
new_h = math.ceil(self.input_size / 3 / 4) # expected to be 11 ro 10
new_w = math.ceil(self.splice / 4) # expected to be 3
inputs = tf.reshape(
inputs, shape=[batch_size * max_time, new_h * new_w * 128])
# Insert linear layer to recude CNN's output demention
# from (new_h * new_w * 128) to 256
with tf.variable_scope('linear') as scope:
inputs = tf.contrib.layers.fully_connected(
inputs=inputs,
num_outputs=256,
activation_fn=tf.nn.relu,
scope=scope)
# Dropout for the VGG-output-hidden connection
inputs = tf.nn.dropout(inputs,
keep_prob_input,
name='dropout_input')
# Reshape back to 3D tensor `[B, T, 256]`
inputs = tf.reshape(inputs, shape=[batch_size, max_time, 256])
initializer = tf.random_uniform_initializer(
minval=-self.parameter_init,
maxval=self.parameter_init)
# Hidden layers
lstm_list = []
with tf.variable_scope('multi_lstm', initializer=initializer) as scope:
for i_layer in range(1, self.num_layers + 1, 1):
if self.lstm_impl == 'BasicLSTMCell':
lstm = tf.contrib.rnn.BasicLSTMCell(
self.num_units,
forget_bias=1.0,
state_is_tuple=True,
activation=tf.tanh)
elif self.lstm_impl == 'LSTMCell':
lstm = tf.contrib.rnn.LSTMCell(
self.num_units,
use_peepholes=self.use_peephole,
cell_clip=self.clip_activation,
num_proj=self.num_proj,
forget_bias=1.0,
state_is_tuple=True)
elif self.lstm_impl == 'LSTMBlockCell':
# NOTE: This should be faster than tf.contrib.rnn.LSTMCell
lstm = tf.contrib.rnn.LSTMBlockCell(
self.num_units,
forget_bias=1.0,
# clip_cell=True,
use_peephole=self.use_peephole)
# TODO: cell clipping (update for rc1.3)
elif self.lstm_impl == 'LSTMBlockFusedCell':
raise NotImplementedError
elif self.lstm_impl == 'CudnnLSTM':
raise NotImplementedError
else:
raise IndexError(
'lstm_impl is "BasicLSTMCell" or "LSTMCell" or ' +
'"LSTMBlockCell" or "LSTMBlockFusedCell" or ' +
'"CudnnLSTM".')
# Dropout for the hidden-hidden connections
lstm = tf.contrib.rnn.DropoutWrapper(
lstm, output_keep_prob=keep_prob_hidden)
lstm_list.append(lstm)
# Stack multiple cells
stacked_lstm = tf.contrib.rnn.MultiRNNCell(
lstm_list, state_is_tuple=True)
# Ignore 2nd return (the last state)
outputs, final_state = tf.nn.dynamic_rnn(
cell=stacked_lstm,
inputs=inputs,
sequence_length=inputs_seq_len,
dtype=tf.float32,
scope=scope)
# NOTE: initial states are zero states by default
if self.return_hidden_states:
return outputs, final_state
# Reshape to apply the same weights over the timesteps
if self.num_proj is None:
outputs = tf.reshape(outputs, shape=[-1, self.num_units])
else:
outputs = tf.reshape(outputs, shape=[-1, self.num_proj])
if self.bottleneck_dim is not None and self.bottleneck_dim != 0:
with tf.variable_scope('bottleneck') as scope:
outputs = tf.contrib.layers.fully_connected(
outputs, self.bottleneck_dim,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
# Dropout for the hidden-output connections
outputs = tf.nn.dropout(
outputs, keep_prob_output, name='dropout_output_bottle')
with tf.variable_scope('output') as scope:
logits_2d = tf.contrib.layers.fully_connected(
outputs, self.num_classes,
activation_fn=None,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
# Reshape back to the original shape
logits = tf.reshape(
logits_2d, shape=[batch_size, -1, self.num_classes])
# Convert to time-major: `[T, B, num_classes]'
logits = tf.transpose(logits, (1, 0, 2))
# Dropout for the hidden-output connections
logits = tf.nn.dropout(
logits, keep_prob_output, name='dropout_output')
return logits, final_state
def __call__(self, inputs, inputs_seq_len, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, T, input_size (num_channels * splice * num_stack * 3)]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states, a tensor of size
`[T, B, num_units (num_proj)]`
final_state: A final hidden state of the encoder
"""
# inputs: 3D tensor `[B, T, input_dim]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B * T, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(inputs,
shape=[
batch_size * max_time, self.num_channels,
self.splice * self.num_stack, 3
])
# NOTE: filter_size: `[H, W, C_in, C_out]`
with tf.variable_scope('VGG1'):
inputs = conv_layer(inputs,
filter_size=[3, 3, 3, 64],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu',
name='conv1')
inputs = conv_layer(inputs,
filter_size=[3, 3, 64, 64],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu',
name='conv2')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[2, 2],
stride=[2, 2],
name='max_pool')
# TODO(hirofumi): try dropout
with tf.variable_scope('VGG2'):
inputs = conv_layer(inputs,
filter_size=[3, 3, 64, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu',
name='conv1')
inputs = conv_layer(inputs,
filter_size=[3, 3, 128, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu',
name='conv2')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[2, 2],
stride=[2, 2],
name='max_pool')
# TODO(hirofumi): try dropout
# Reshape to 2D tensor `[B * T, new_h * new_w * C_out]`
new_h = math.ceil(self.num_channels / 4)
new_w = math.ceil(self.splice * self.num_stack / 4)
channel_out = inputs.shape.as_list()[-1]
inputs = tf.reshape(
inputs, shape=[batch_size * max_time, new_h * new_w * channel_out])
# Insert linear layer to recude CNN's output demention
# from (new_h * new_w * C_out) to 256
with tf.variable_scope('bridge') as scope:
inputs = tf.contrib.layers.fully_connected(
inputs=inputs,
num_outputs=256,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
# Dropout for the VGG-output-hidden connection
inputs = tf.nn.dropout(inputs, keep_prob, name='dropout_pipe')
# Reshape back to 3D tensor `[B, T, 256]`
inputs = tf.reshape(inputs, shape=[batch_size, max_time, 256])
initializer = tf.random_uniform_initializer(
minval=-self.parameter_init, maxval=self.parameter_init)
if self.lstm_impl == 'BasicLSTMCell':
outputs, final_state = basiclstmcell(self.num_units,
self.num_layers, inputs,
inputs_seq_len, keep_prob,
initializer, self.time_major)
elif self.lstm_impl == 'LSTMCell':
outputs, final_state = lstmcell(self.num_units, self.num_proj,
self.num_layers, self.use_peephole,
self.clip_activation, inputs,
inputs_seq_len, keep_prob,
initializer, self.time_major)
elif self.lstm_impl == 'LSTMBlockCell':
outputs, final_state = lstmblockcell(self.num_units,
self.num_layers,
self.use_peephole,
self.clip_activation, inputs,
inputs_seq_len, keep_prob,
initializer, self.time_major)
elif self.lstm_impl == 'LSTMBlockFusedCell':
outputs, final_state = lstmblockfusedcell(self.num_units,
self.num_layers, inputs,
inputs_seq_len,
keep_prob, initializer,
self.time_major)
elif self.lstm_impl == 'CudnnLSTM':
outputs, final_state = cudnnlstm(self.num_units, self.num_layers,
inputs, inputs_seq_len, keep_prob,
initializer, self.time_major)
else:
raise IndexError('lstm_impl is "BasicLSTMCell" or "LSTMCell" or ' +
'"LSTMBlockCell" or "LSTMBlockFusedCell" or ' +
'"CudnnLSTM".')
return outputs, final_state
def __call__(self, inputs, inputs_seq_len, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, T, input_size (num_channels * (splice * num_stack) * 3)]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size
`[T, B, num_units (num_proj)]`
otherwise, `[B, T, num_units (num_proj)]`
final_state: A final hidden state of the encoder
"""
# inputs: 3D tensor `[B, T, input_dim]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
# For debug
# print(input_dim)
# print(self.num_channels)
# print(self.splice)
# print(self.num_stack)
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B * T, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(inputs,
shape=[
batch_size * max_time, self.num_channels,
self.splice * self.num_stack, 3
])
# NOTE: filter_size: `[H, W, C_in, C_out]`
with tf.variable_scope('CNN1'):
inputs = conv_layer(inputs,
filter_size=[11, 21, 3, 32],
stride=[3, 2],
parameter_init=self.parameter_init,
activation='relu')
# inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
with tf.variable_scope('CNN2'):
inputs = conv_layer(inputs,
filter_size=[11, 11, 32, 32],
stride=[1, 2],
parameter_init=self.parameter_init,
activation='relu')
# inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
with tf.variable_scope('CNN3'):
inputs = conv_layer(inputs,
filter_size=[3, 3, 32, 96],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
# inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
inputs = tf.nn.dropout(inputs, keep_prob)
# Reshape to 3D tensor `[B, T, new_h * new_w * C_out]`
inputs = tf.reshape(
inputs,
shape=[batch_size, max_time,
np.prod(inputs.shape.as_list()[-3:])])
initializer = tf.random_uniform_initializer(
minval=-self.parameter_init, maxval=self.parameter_init)
if self.lstm_impl == 'BasicLSTMCell':
outputs, final_state = basiclstmcell(self.num_units,
self.num_layers, inputs,
inputs_seq_len, keep_prob,
initializer, self.time_major)
elif self.lstm_impl == 'LSTMCell':
outputs, final_state = lstmcell(self.num_units, self.num_proj,
self.num_layers, self.use_peephole,
self.clip_activation, inputs,
inputs_seq_len, keep_prob,
initializer, self.time_major)
elif self.lstm_impl == 'LSTMBlockCell':
outputs, final_state = lstmblockcell(self.num_units,
self.num_layers,
self.use_peephole,
self.clip_activation, inputs,
inputs_seq_len, keep_prob,
initializer, self.time_major)
elif self.lstm_impl == 'LSTMBlockFusedCell':
outputs, final_state = lstmblockfusedcell(self.num_units,
self.num_layers, inputs,
inputs_seq_len,
keep_prob, initializer,
self.time_major)
elif self.lstm_impl == 'CudnnLSTM':
outputs, final_state = cudnnlstm(self.num_units, self.num_layers,
inputs, inputs_seq_len, keep_prob,
initializer, self.time_major)
else:
raise IndexError('lstm_impl is "BasicLSTMCell" or "LSTMCell" or ' +
'"LSTMBlockCell" or "LSTMBlockFusedCell" or ' +
'"CudnnLSTM".')
# Reshape to 2D tensor `[B * T (T * B), output_dim]`
output_dim = outputs.shape.as_list()[-1]
outputs = tf.reshape(outputs,
shape=[batch_size * max_time, output_dim])
with tf.variable_scope('fc1') as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=896,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
outputs = tf.nn.dropout(outputs, keep_prob)
with tf.variable_scope('fc2') as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=74,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
output_dim = outputs.shape.as_list()[-1]
if self.time_major:
# Reshape back to 3D tensor `[T, B, 74]`
outputs = tf.reshape(outputs,
shape=[max_time, batch_size, output_dim])
else:
# Reshape back to 3D tensor `[B, T, 74]`
outputs = tf.reshape(outputs,
shape=[batch_size, max_time, output_dim])
return outputs, final_state
def __call__(self, inputs, inputs_seq_len, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, T, input_size (num_channels * splice * num_stack * 3)]`
inputs_seq_len (placeholder): A tensor of size` [B]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size `[T, B, output_dim]`
otherwise, `[B, T, output_dim]`
final_state: None
"""
# inputs: 3D tensor `[B, T, input_dim]`
batch_size = tf.shape(inputs)[0]
max_time = tf.shape(inputs)[1]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B * T, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(inputs,
shape=[
batch_size * max_time, self.num_channels,
self.splice * self.num_stack, 3
])
# Choose the activation function
activation = 'relu'
# activation = 'prelu'
# activation = 'maxout'
# TODO: add prelu and maxout layers
# 1-4th layers
with tf.variable_scope('CNN1'):
for i_layer in range(1, 5, 1):
if i_layer == 1:
inputs = conv_layer(inputs,
filter_size=[3, 5, 3, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation=activation,
name='conv1')
inputs = max_pool(inputs,
pooling_size=[3, 1],
stride=[3, 1],
name='pool')
else:
inputs = conv_layer(inputs,
filter_size=[3, 5, 128, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation=activation,
name='conv%d' % i_layer)
# NOTE: No poling
inputs = batch_normalization(inputs, is_training=is_training)
# inputs = tf.nn.dropout(inputs, keep_prob)
# TODO: try Weight decay
# 5-10th layers
with tf.variable_scope('CNN2'):
for i_layer in range(5, 11, 1):
if i_layer == 5:
inputs = conv_layer(inputs,
filter_size=[3, 5, 128, 256],
stride=[1, 1],
parameter_init=self.parameter_init,
activation=activation,
name='conv1')
# NOTE: No poling
else:
inputs = conv_layer(inputs,
filter_size=[3, 5, 256, 256],
stride=[1, 1],
parameter_init=self.parameter_init,
activation=activation,
name='conv%d' % i_layer)
# NOTE: No poling
inputs = batch_normalization(inputs, is_training=is_training)
# inputs = tf.nn.dropout(inputs, keep_prob)
# TODO: try Weight decay
# Reshape to 2D tensor `[B * T, new_h * new_w * C_out]`
new_h = math.ceil(self.num_channels / 3)
new_w = self.splice * self.num_stack
channel_out = inputs.shape.as_list()[-1]
outputs = tf.reshape(
inputs, shape=[batch_size * max_time, new_h * new_w * channel_out])
# 11-14th layers
for i in range(1, 4, 1):
with tf.variable_scope('fc%d' % i) as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=1024,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
outputs = batch_normalization(outputs, is_training=is_training)
outputs = tf.nn.dropout(outputs, keep_prob)
# TODO: try Weight decay
# Reshape back to 3D tensor `[B, T, 1024]`
logits = tf.reshape(outputs, shape=[batch_size, max_time, 1024])
if self.time_major:
# Convert to time-major: `[T, B, 1024]'
logits = tf.transpose(logits, [1, 0, 2])
return logits, None
def __call__(self, inputs, keep_prob, is_training):
"""Construct model graph.
Args:
inputs (placeholder): A tensor of size
`[B, input_size (num_channels * splice * num_stack * 3)]`
keep_prob (placeholder, float): A probability to keep nodes
in the hidden-hidden connection
is_training (bool):
Returns:
outputs: Encoder states.
if time_major is True, a tensor of size `[T, B, output_dim]`
otherwise, `[B, output_dim]`
"""
# inputs: 2D tensor `[B, input_dim]`
batch_size = tf.shape(inputs)[0]
input_dim = inputs.shape.as_list()[-1]
# NOTE: input_dim: num_channels * splice * num_stack * 3
# for debug
# print(input_dim) # 1200
# print(self.num_channels) # 40
# print(self.splice) # 5
# print(self.num_stack) # 2
assert input_dim == self.num_channels * self.splice * self.num_stack * 3
# Reshape to 4D tensor `[B, num_channels, splice * num_stack, 3]`
inputs = tf.reshape(
inputs,
shape=[batch_size, self.num_channels, self.splice * self.num_stack, 3])
# NOTE: filter_size: `[H, W, C_in, C_out]`
with tf.variable_scope('CNN1'):
inputs = conv_layer(inputs,
filter_size=[9, 9, 3, 128],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[3, 1],
stride=[3, 1],
name='max_pool')
with tf.variable_scope('CNN2'):
inputs = conv_layer(inputs,
filter_size=[3, 4, 128, 256],
stride=[1, 1],
parameter_init=self.parameter_init,
activation='relu')
inputs = batch_normalization(inputs, is_training=is_training)
inputs = max_pool(inputs,
pooling_size=[1, 1],
stride=[1, 1],
name='max_pool')
# Reshape to 2D tensor `[B, new_h * new_w * C_out]`
outputs = tf.reshape(
inputs, shape=[batch_size, np.prod(inputs.shape.as_list()[-3:])])
for i in range(1, 5, 1):
with tf.variable_scope('fc%d' % (i)) as scope:
outputs = tf.contrib.layers.fully_connected(
inputs=outputs,
num_outputs=2048,
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(
stddev=self.parameter_init),
biases_initializer=tf.zeros_initializer(),
scope=scope)
return outputs