def get_constants(self, inputs, training=None):
constants = []
if self.implementation != 0 and 0 < self.dropout < 1:
input_shape = K.int_shape(inputs)
input_dim = input_shape[-1]
ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
ones = K.tile(ones, (1, int(input_dim)))
def dropped_inputs():
return K.dropout(ones, self.dropout)
dp_mask = [
K.in_train_phase(dropped_inputs, ones, training=training)
for _ in range(3)
]
constants.append(dp_mask)
else:
constants.append([K.cast_to_floatx(1.) for _ in range(3)])
if 0 < self.recurrent_dropout < 1:
ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
ones = K.tile(ones, (1, self.units))
def dropped_inputs(): # pylint: disable=function-redefined
return K.dropout(ones, self.recurrent_dropout)
rec_dp_mask = [
K.in_train_phase(dropped_inputs, ones, training=training)
for _ in range(3)
]
constants.append(rec_dp_mask)
else:
constants.append([K.cast_to_floatx(1.) for _ in range(3)])
return constants
def call(self, inputs):
if not isinstance(inputs, list):
raise ValueError('A merge layer should be called ' 'on a list of inputs.')
if self._reshape_required:
reshaped_inputs = []
input_ndims = list(map(K.ndim, inputs))
if None not in input_ndims:
# If ranks of all inputs are available,
# we simply expand each of them at axis=1
# until all of them have the same rank.
max_ndim = max(input_ndims)
for x in inputs:
x_ndim = K.ndim(x)
for _ in range(max_ndim - x_ndim):
x = K.expand_dims(x, 1)
reshaped_inputs.append(x)
return self._merge_function(reshaped_inputs)
else:
# Transpose all inputs so that batch size is the last dimension.
# (batch_size, dim1, dim2, ... ) -> (dim1, dim2, ... , batch_size)
transposed = False
for x in inputs:
x_ndim = K.ndim(x)
if x_ndim is None:
x_shape = K.shape(x)
batch_size = x_shape[0]
new_shape = K.concatenate([x_shape[1:], K.expand_dims(batch_size)])
x_transposed = K.reshape(x,
K.stack([batch_size,
K.prod(x_shape[1:])]))
x_transposed = K.permute_dimensions(x_transposed, (1, 0))
x_transposed = K.reshape(x_transposed, new_shape)
reshaped_inputs.append(x_transposed)
transposed = True
elif x_ndim > 1:
dims = list(range(1, x_ndim)) + [0]
reshaped_inputs.append(K.permute_dimensions(x, dims))
transposed = True
else:
# We don't transpose inputs if they are 1D vectors or scalars.
reshaped_inputs.append(x)
y = self._merge_function(reshaped_inputs)
y_ndim = K.ndim(y)
if transposed:
# If inputs have been transposed, we have to transpose the output too.
if y_ndim is None:
y_shape = K.shape(y)
y_ndim = K.shape(y_shape)[0]
batch_size = y_shape[y_ndim - 1]
new_shape = K.concatenate(
[K.expand_dims(batch_size), y_shape[:y_ndim - 1]])
y = K.reshape(y, (-1, batch_size))
y = K.permute_dimensions(y, (1, 0))
y = K.reshape(y, new_shape)
elif y_ndim > 1:
dims = [y_ndim - 1] + list(range(y_ndim - 1))
y = K.permute_dimensions(y, dims)
return y
else:
return self._merge_function(inputs)
def _linear(self, x, kernel, bias):
in_shape = x.shape
if len(in_shape) > 2:
x_shape = [int(dim) for dim in x.shape]
x = K.reshape(x, (x_shape[0] * x_shape[1], x_shape[2]))
x = K.dot(x, kernel)
x = K.bias_add(x, bias)
if len(in_shape) > 2:
x = K.reshape(x, (x_shape[0], x_shape[1], int(kernel.shape[1])))
return x
def _layer_norm(self, x, offset, scale):
in_shape = x.shape
if len(in_shape) > 2:
x_shape = [int(dim) for dim in x.shape]
x = K.reshape(x, (x_shape[0] * x_shape[1], x_shape[2]))
mean, var = tf.nn.moments(x, [1], keep_dims=True)
x = tf.nn.batch_normalization(x, mean, var, offset, scale, K.epsilon())
if len(in_shape) > 2:
x = K.reshape(x, x_shape)
return x
def call(self, inputs, training=None, mask=None):
kwargs = {}
if has_arg(self.layer.call, 'training'):
kwargs['training'] = training
uses_learning_phase = False # pylint: disable=redefined-outer-name
input_shape = K.int_shape(inputs)
if input_shape[0]:
# batch size matters, use rnn-based implementation
def step(x, _):
global uses_learning_phase # pylint: disable=global-variable-undefined
output = self.layer.call(x, **kwargs)
if hasattr(output, '_uses_learning_phase'):
uses_learning_phase = (output._uses_learning_phase or
uses_learning_phase)
return output, []
_, outputs, _ = K.rnn(
step,
inputs,
initial_states=[],
unroll=False)
y = outputs
else:
# No batch size specified, therefore the layer will be able
# to process batches of any size.
# We can go with reshape-based implementation for performance.
input_length = input_shape[1]
if not input_length:
input_length = K.shape(inputs)[1]
# Shape: (num_samples * timesteps, ...). And track the
# transformation in self._input_map.
input_uid = tf_layers_util.object_list_uid(inputs)
inputs = K.reshape(inputs, (-1,) + input_shape[2:])
self._input_map[input_uid] = inputs
# (num_samples * timesteps, ...)
y = self.layer.call(inputs, **kwargs)
if hasattr(y, '_uses_learning_phase'):
uses_learning_phase = y._uses_learning_phase
# Shape: (num_samples, timesteps, ...)
output_shape = self.compute_output_shape(input_shape).as_list()
y = K.reshape(y, (-1, input_length) + tuple(output_shape[2:]))
# Apply activity regularizer if any:
if (hasattr(self.layer, 'activity_regularizer') and
self.layer.activity_regularizer is not None):
regularization_loss = self.layer.activity_regularizer(y)
self.add_loss(regularization_loss, inputs)
if uses_learning_phase:
y._uses_learning_phase = True
return y
def call(self, inputs, training=None, mask=None):
kwargs = {}
if has_arg(self.layer.call, 'training'):
kwargs['training'] = training
uses_learning_phase = False # pylint: disable=redefined-outer-name
input_shape = K.int_shape(inputs)
if input_shape[0]:
# batch size matters, use rnn-based implementation
def step(x, _):
global uses_learning_phase # pylint: disable=global-variable-undefined
output = self.layer.call(x, **kwargs)
if hasattr(output, '_uses_learning_phase'):
uses_learning_phase = (output._uses_learning_phase
or uses_learning_phase)
return output, []
_, outputs, _ = K.rnn(step,
inputs,
initial_states=[],
unroll=False)
y = outputs
else:
# No batch size specified, therefore the layer will be able
# to process batches of any size.
# We can go with reshape-based implementation for performance.
input_length = input_shape[1]
if not input_length:
input_length = K.shape(inputs)[1]
# Shape: (num_samples * timesteps, ...). And track the
# transformation in self._input_map.
input_uid = tf_layers_util.object_list_uid(inputs)
inputs = K.reshape(inputs, (-1, ) + input_shape[2:])
self._input_map[input_uid] = inputs
# (num_samples * timesteps, ...)
y = self.layer.call(inputs, **kwargs)
if hasattr(y, '_uses_learning_phase'):
uses_learning_phase = y._uses_learning_phase
# Shape: (num_samples, timesteps, ...)
output_shape = self.compute_output_shape(input_shape).as_list()
y = K.reshape(y, (-1, input_length) + tuple(output_shape[2:]))
# Apply activity regularizer if any:
if (hasattr(self.layer, 'activity_regularizer')
and self.layer.activity_regularizer is not None):
regularization_loss = self.layer.activity_regularizer(y)
self.add_loss(regularization_loss, inputs)
if uses_learning_phase:
y._uses_learning_phase = True
return y
def _time_distributed_dense(x,
w,
b=None,
dropout=None,
input_dim=None,
output_dim=None,
timesteps=None,
training=None):
"""Apply `y . w + b` for every temporal slice y of x.
Arguments:
x: input tensor.
w: weight matrix.
b: optional bias vector.
dropout: whether to apply dropout (same dropout mask
for every temporal slice of the input).
input_dim: integer; optional dimensionality of the input.
output_dim: integer; optional dimensionality of the output.
timesteps: integer; optional number of timesteps.
training: training phase tensor or boolean.
Returns:
Output tensor.
"""
if not input_dim:
input_dim = K.shape(x)[2]
if not timesteps:
timesteps = K.shape(x)[1]
if not output_dim:
output_dim = K.shape(w)[1]
if dropout is not None and 0. < dropout < 1.:
# apply the same dropout pattern at every timestep
ones = K.ones_like(K.reshape(x[:, 0, :], (-1, input_dim)))
dropout_matrix = K.dropout(ones, dropout)
expanded_dropout_matrix = K.repeat(dropout_matrix, timesteps)
x = K.in_train_phase(x * expanded_dropout_matrix, x, training=training)
# collapse time dimension and batch dimension together
x = K.reshape(x, (-1, input_dim))
x = K.dot(x, w)
if b is not None:
x = K.bias_add(x, b)
# reshape to 3D tensor
if K.backend() == 'tensorflow':
x = K.reshape(x, K.stack([-1, timesteps, output_dim]))
x.set_shape([None, None, output_dim])
else:
x = K.reshape(x, (-1, timesteps, output_dim))
return x
def call(self, inputs):
# In case the target shape is not fully defined,
# we need access to the shape of x.
target_shape = self.target_shape
if -1 in target_shape:
# target shape not fully defined
target_shape = self._compute_output_shape(inputs.get_shape())
target_shape = target_shape.as_list()[1:]
return K.reshape(inputs, (-1,) + tuple(target_shape))
def _multihead_attention(self, memory):
"""Perform multi-head attention from 'Attention is All You Need'.
Implementation of the attention mechanism from
https://arxiv.org/abs/1706.03762.
Args:
memory: Memory tensor to perform attention on.
Returns:
new_memory: New memory tensor.
"""
batch_size = int(memory.shape[0])
qkv = self._linear(memory, self.kernel_qkv, self.bias_qkv)
# qkv = self._layer_norm(qkv, self.offset_qkv, self.scale_qkv)
mem_slots = memory.get_shape().as_list()[1] # Denoted as N.
# [B, N, F] -> [B, N, H, F/H]
qkv_reshape = K.reshape(
qkv, (batch_size, mem_slots, self.num_heads, self.qkv_size))
# [B, N, H, F/H] -> [B, H, N, F/H]
qkv_transpose = K.permute_dimensions(qkv_reshape, [0, 2, 1, 3])
q, k, v = tf.split(qkv_transpose,
[self.key_size, self.key_size, self.value_size], -1)
q *= self.qkv_size**-0.5
dot_product = tf.matmul(q, k, transpose_b=True) # [B, H, N, N]
weights = K.softmax(dot_product)
output = tf.matmul(weights, v) # [B, H, N, V]
# [B, H, N, V] -> [B, N, H, V]
output_transpose = K.permute_dimensions(output, [0, 2, 1, 3])
# [B, N, H, V] -> [B, N, H * V]
new_memory = K.reshape(output_transpose,
(batch_size, mem_slots, self.mem_size))
return new_memory
def compute_spectral_normal(self, training=True):
# Spectrally Normalized Weight
if self.spectral_normalization:
# Get kernel tensor shape
W_shape = self.kernel.shape.as_list()
# Flatten the Tensor
W_mat = K.reshape(self.kernel, [W_shape[-1], -1]) # [out_c, N]
sigma, u, v = power_iteration(W_mat, self.u)
if training:
# Update estimated 1st singular vector
self.u.assign(u)
return self.kernel / sigma
else:
return self.kernel
def call(self, inputs, memory, training=None):
batch_size = int(inputs.shape[0])
memory = K.reshape(memory, (batch_size, self.mem_slots, self.mem_size))
inputs = self._linear(inputs, self.kernel_in, self.bias_in)
inputs_reshape = K.expand_dims(inputs, axis=1)
memory_plus_input = K.concatenate([memory, inputs_reshape], axis=1)
next_memory = self._attend_over_memory(memory_plus_input)
n = inputs_reshape.get_shape().as_list()[1]
next_memory = next_memory[:, :-n, :]
input_gate, forget_gate = self._create_gates(inputs_reshape, memory)
next_memory = input_gate * K.tanh(next_memory)
next_memory += forget_gate * memory
next_memory = K.batch_flatten(next_memory)
return next_memory, (next_memory, )
def call(self, inputs):
if not isinstance(inputs, list):
raise ValueError('A merge layer should be called '
'on a list of inputs.')
if self._reshape_required:
reshaped_inputs = []
input_ndims = list(map(K.ndim, inputs))
if None not in input_ndims:
# If ranks of all inputs are available,
# we simply expand each of them at axis=1
# until all of them have the same rank.
max_ndim = max(input_ndims)
for x in inputs:
x_ndim = K.ndim(x)
for _ in range(max_ndim - x_ndim):
x = K.expand_dims(x, 1)
reshaped_inputs.append(x)
return self._merge_function(reshaped_inputs)
else:
# Transpose all inputs so that batch size is the last dimension.
# (batch_size, dim1, dim2, ... ) -> (dim1, dim2, ... , batch_size)
transposed = False
for x in inputs:
x_ndim = K.ndim(x)
if x_ndim is None:
x_shape = K.shape(x)
batch_size = x_shape[0]
new_shape = K.concatenate(
[x_shape[1:],
K.expand_dims(batch_size)])
x_transposed = K.reshape(
x, K.stack([batch_size,
K.prod(x_shape[1:])]))
x_transposed = K.permute_dimensions(
x_transposed, (1, 0))
x_transposed = K.reshape(x_transposed, new_shape)
reshaped_inputs.append(x_transposed)
transposed = True
elif x_ndim > 1:
dims = list(range(1, x_ndim)) + [0]
reshaped_inputs.append(K.permute_dimensions(x, dims))
transposed = True
else:
# We don't transpose inputs if they are 1D vectors or scalars.
reshaped_inputs.append(x)
y = self._merge_function(reshaped_inputs)
y_ndim = K.ndim(y)
if transposed:
# If inputs have been transposed, we have to transpose the output too.
if y_ndim is None:
y_shape = K.shape(y)
y_ndim = K.shape(y_shape)[0]
batch_size = y_shape[y_ndim - 1]
new_shape = K.concatenate(
[K.expand_dims(batch_size), y_shape[:y_ndim - 1]])
y = K.reshape(y, (-1, batch_size))
y = K.permute_dimensions(y, (1, 0))
y = K.reshape(y, new_shape)
elif y_ndim > 1:
dims = [y_ndim - 1] + list(range(y_ndim - 1))
y = K.permute_dimensions(y, dims)
return y
else:
return self._merge_function(inputs)
def call(self, inputs):
return K.reshape(inputs, (K.shape(inputs)[0], ) + self.target_shape)
def call(self, inputs): return K.reshape(inputs, (K.shape(inputs)[0],) + self.target_shape)
