def __init__(self,
filters,
kernel_size,
strides=(1, 1),
padding='valid',
data_format=None,
dilation_rate=(1, 1),
activation=None,
use_bias=True,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs):
super(Conv2D, self).__init__(
filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=getters.get_activation(activation) if activation else activation,
use_bias=use_bias,
kernel_initializer=getters.get_initializer(kernel_initializer),
bias_initializer=getters.get_initializer(bias_initializer),
kernel_regularizer=getters.get_regularizer(kernel_regularizer),
bias_regularizer=getters.get_regularizer(bias_regularizer),
activity_regularizer=getters.get_regularizer(activity_regularizer),
kernel_constraint=getters.get_constraint(kernel_constraint),
bias_constraint=getters.get_constraint(bias_constraint),
**kwargs)
def __init__(self,
filters,
kernel_size,
strides=(1, 1),
padding='valid',
data_format=None,
depth_multiplier=1,
activation=None,
use_bias=True,
depthwise_initializer='glorot_uniform',
pointwise_initializer='glorot_uniform',
bias_initializer='zeros',
depthwise_regularizer=None,
pointwise_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
depthwise_constraint=None,
pointwise_constraint=None,
bias_constraint=None,
**kwargs):
super(SeparableConv2D, self).__init__(
filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
depth_multiplier=depth_multiplier,
activation=getters.get_activation(activation),
use_bias=use_bias,
depthwise_initializer=getters.get_initializer(
depthwise_initializer),
pointwise_initializer=getters.get_initializer(
pointwise_initializer),
bias_initializer=getters.get_initializer(bias_initializer),
depthwise_regularizer=getters.get_regularizer(
depthwise_regularizer),
pointwise_regularizer=getters.get_regularizer(
pointwise_regularizer),
bias_regularizer=getters.get_regularizer(bias_regularizer),
activity_regularizer=getters.get_regularizer(activity_regularizer),
depthwise_constraint=getters.get_constraint(depthwise_constraint),
pointwise_constraint=getters.get_constraint(pointwise_constraint),
bias_constraint=getters.get_constraint(bias_constraint),
**kwargs)
def _build(self, incoming, *args, **kwargs):
"""
Args:
incoming: (2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Returns:
2D Tensor [samples, num_units].
"""
self._declare_dependencies()
input_shape = get_shape(incoming)
incoming = validate_dtype(incoming)
assert len(
input_shape) > 1, 'Incoming Tensor shape must be at least 2-D'
n_inputs = total_tensor_depth(tensor_shape=input_shape)
regularizer = getters.get_regularizer(self.regularizer,
scale=self.scale,
collect=True)
self._w = variable(name='w',
shape=[n_inputs, self.num_units],
dtype=incoming.dtype,
regularizer=regularizer,
initializer=getters.get_initializer(
self.weights_init),
trainable=self.trainable,
restore=self.restore)
track(self._w, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 2:
inference = tf.reshape(tensor=inference, shape=[-1, n_inputs])
inference = tf.matmul(a=inference, b=self._w)
self._b = None
if self.bias:
self._b = variable(name='b',
shape=[self.num_units],
dtype=incoming.dtype,
initializer=getters.get_initializer(
self.bias_init),
trainable=self.trainable,
restore=self.restore)
track(self._b, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
inference = tf.nn.bias_add(value=inference, bias=self._b)
if self.activation:
inference = getters.get_activation(self.activation,
collect=True)(inference)
if self._dropout:
inference = self._dropout(inference)
track(inference, tf.GraphKeys.LAYER_TENSOR, self.module_name)
return inference
def __init__(self,
input_dim,
output_dim,
embeddings_initializer='uniform',
embeddings_regularizer=None,
activity_regularizer=None,
embeddings_constraint=None,
mask_zero=False,
input_length=None,
**kwargs):
super(Embedding, self).__init__(
input_dim=input_dim,
output_dim=output_dim,
embeddings_initializer=getters.get_initializer(embeddings_initializer),
embeddings_regularizer=getters.get_regularizer(embeddings_regularizer),
activity_regularizer=getters.get_regularizer(activity_regularizer),
embeddings_constraint=getters.get_constraint(embeddings_constraint),
mask_zero=mask_zero,
input_length=input_length,
**kwargs)
def __init__(self,
units,
activation='tanh',
recurrent_activation='hard_sigmoid',
use_bias=True,
kernel_initializer='glorot_uniform',
recurrent_initializer='orthogonal',
bias_initializer='zeros',
unit_forget_bias=True,
kernel_regularizer=None,
recurrent_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
recurrent_constraint=None,
bias_constraint=None,
dropout=0.,
recurrent_dropout=0.,
**kwargs):
super(LSTM, self).__init__(
units=units,
activation=getters.get_activation(activation),
recurrent_activation=getters.get_activation(recurrent_activation),
use_bias=use_bias,
kernel_initializer=getters.get_initializer(kernel_initializer),
recurrent_initializer=getters.get_initializer(
recurrent_initializer),
bias_initializer=getters.get_initializer(bias_initializer),
unit_forget_bias=unit_forget_bias,
kernel_regularizer=getters.get_regularizer(kernel_regularizer),
recurrent_regularizer=getters.get_regularizer(
recurrent_regularizer),
bias_regularizer=getters.get_regularizer(bias_regularizer),
activity_regularizer=getters.get_regularizer(activity_regularizer),
kernel_constraint=getters.get_constraint(kernel_constraint),
recurrent_constraint=getters.get_constraint(recurrent_constraint),
bias_constraint=getters.get_constraint(bias_constraint),
dropout=dropout,
recurrent_dropout=recurrent_dropout,
**kwargs)
def _build(self, incoming, *args, **kwargs):
"""
Args:
incoming: (2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Returns:
2D Tensor [samples, num_units].
"""
self._declare_dependencies()
input_shape = get_shape(incoming)
assert len(input_shape) > 1, 'Incoming Tensor shape must be at least 2-D'
n_inputs = total_tensor_depth(tensor_shape=input_shape)
regularizer = getters.get_regularizer(self.regularizer, scale=self.scale, collect=True)
initializer = getters.get_initializer(self.weights_init)
self._w = variable(name='w', shape=[n_inputs, self.num_units], regularizer=regularizer,
initializer=initializer, trainable=self.trainable,
restore=self.restore)
track(self._w, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
self._b = variable(name='b', shape=[self.num_units],
initializer=getters.get_initializer(self.bias_init),
trainable=self.trainable, restore=self.restore)
track(self._b, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
# Weight and bias for the transform gate
self._w_t = variable(name='w_t', shape=[n_inputs, self.num_units],
regularizer=None, initializer=initializer,
trainable=self.trainable, restore=self.restore)
track(self._w_t, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
self._b_t = variable(name='b_t', shape=[self.num_units],
initializer=tf.constant_initializer(-1),
trainable=self.trainable, restore=self.restore)
track(self._b_t, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
# If input is not 2d, flatten it.
if len(input_shape) > 2:
incoming = tf.reshape(tensor=incoming, shape=[-1, n_inputs])
H = getters.get_activation(self.activation)(tf.matmul(a=incoming, b=self._w) + self._b)
T = tf.sigmoid(tf.matmul(a=incoming, b=self._w_t) + self._b_t)
if self._transform_dropout:
T = self._transform_dropout(T)
C = tf.subtract(x=1.0, y=T)
inference = tf.add(x=tf.multiply(x=H, y=T), y=tf.multiply(x=incoming, y=C))
track(inference, tf.GraphKeys.ACTIVATIONS)
track(inference, tf.GraphKeys.LAYER_TENSOR, self.module_name)
return inference
def variable(name,
shape=None,
dtype=tf.float32,
initializer=None,
regularizer=None,
trainable=True,
collections=None,
device='',
restore=True):
"""Instantiate a new variable.
Args:
name: `str`. A name for this variable.
shape: list of `int`. The variable shape (optional).
dtype: `type`. The variable data type.
initializer: `str` or `Tensor`. The variable initialization.
regularizer: `str` or `Tensor`. The variable regularizer.
trainable: `bool`. If True, this variable weights will be trained.
collections: `str`. A collection to add the new variable to (optional).
device: `str`. Device ID to store the variable. Default: '/cpu:0'.
restore: `bool`. Restore or not this variable when loading a pre-trained model.
Returns:
A Variable.
"""
if isinstance(initializer, six.string_types):
initializer = getters.get_initializer(initializer)
# Remove shape param if initializer is a Tensor
if not callable(initializer) and isinstance(initializer, tf.Tensor):
shape = None
if isinstance(regularizer, six.string_types):
regularizer = getters.get_regularizer(regularizer)
with tf.device(device_name_or_function=device):
var = tf.get_variable(name=name,
shape=shape,
dtype=dtype,
initializer=initializer,
regularizer=regularizer,
trainable=trainable,
collections=collections)
if not restore:
# TODO adapt restoring saver
tf.add_to_collection(name=tf.GraphKeys.EXCL_RESTORE_VARIABLES,
value=var)
return var
def __init__(self,
filters,
kernel_size,
strides=(1, 1),
padding='valid',
data_format=None,
dilation_rate=(1, 1),
activation='tanh',
recurrent_activation='hard_sigmoid',
use_bias=True,
kernel_initializer='glorot_uniform',
recurrent_initializer='orthogonal',
bias_initializer='zeros',
unit_forget_bias=True,
kernel_regularizer=None,
recurrent_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
recurrent_constraint=None,
bias_constraint=None,
return_sequences=False,
go_backwards=False,
stateful=False,
dropout=0.,
recurrent_dropout=0.,
**kwargs):
super(ConvLSTM2D, self).__init__(
filters,
kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=getters.get_activation(activation),
recurrent_activation=getters.get_activation(recurrent_activation),
use_bias=use_bias,
kernel_initializer=getters.get_initializer(kernel_initializer),
recurrent_initializer=getters.get_initializer(
recurrent_initializer),
bias_initializer=getters.get_initializer(bias_initializer),
unit_forget_bias=unit_forget_bias,
kernel_regularizer=getters.get_regularizer(kernel_regularizer),
recurrent_regularizer=getters.get_regularizer(
recurrent_regularizer),
bias_regularizer=getters.get_regularizer(bias_regularizer),
activity_regularizer=getters.get_regularizer(activity_regularizer),
kernel_constraint=getters.get_regularizer(kernel_constraint),
recurrent_constraint=getters.get_regularizer(recurrent_constraint),
bias_constraint=getters.get_constraint(bias_constraint),
return_sequences=return_sequences,
go_backwards=go_backwards,
stateful=stateful,
dropout=dropout,
recurrent_dropout=recurrent_dropout,
**kwargs)
def __init__(self,
units,
activation=None,
use_bias=True,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs):
super(Dense, self).__init__(
units,
activation=getters.get_activation(activation) if activation else activation,
use_bias=use_bias,
kernel_initializer=getters.get_initializer(kernel_initializer),
bias_initializer=getters.get_initializer(bias_initializer),
kernel_regularizer=getters.get_regularizer(kernel_regularizer),
bias_regularizer=getters.get_regularizer(bias_regularizer),
activity_regularizer=getters.get_regularizer(activity_regularizer),
kernel_constraint=getters.get_constraint(kernel_constraint),
bias_constraint=getters.get_constraint(bias_constraint),
**kwargs)
def _build(self, incoming, *args, **kwargs):
"""
Args:
incoming: (2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Returns:
2D Tensor [samples, num_units].
"""
self._declare_dependencies()
input_shape = get_shape(incoming)
incoming = validate_dtype(incoming)
assert len(input_shape) > 1, 'Incoming Tensor shape must be at least 2-D'
n_inputs = total_tensor_depth(tensor_shape=input_shape)
regularizer = getters.get_regularizer(self.regularizer, scale=self.scale, collect=True)
self._w = variable(
name='w', shape=[n_inputs, self.num_units], dtype=incoming.dtype, regularizer=regularizer,
initializer=getters.get_initializer(self.weights_init), trainable=self.trainable,
restore=self.restore)
track(self._w, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 2:
inference = tf.reshape(tensor=inference, shape=[-1, n_inputs])
inference = tf.matmul(a=inference, b=self._w)
self._b = None
if self.bias:
self._b = variable(name='b', shape=[self.num_units], dtype=incoming.dtype,
initializer=getters.get_initializer(self.bias_init),
trainable=self.trainable, restore=self.restore)
track(self._b, tf.GraphKeys.LAYER_VARIABLES, self.module_name)
inference = tf.nn.bias_add(value=inference, bias=self._b)
if self.activation:
inference = getters.get_activation(self.activation, collect=True)(inference)
if self._dropout:
inference = self._dropout(inference)
track(inference, tf.GraphKeys.LAYER_TENSOR, self.module_name)
return inference
def variable(name, shape=None, dtype=tf.float32, initializer=None, regularizer=None,
trainable=True, collections=None, device='', restore=True):
"""Instantiate a new variable.
Args:
name: `str`. A name for this variable.
shape: list of `int`. The variable shape (optional).
dtype: `type`. The variable data type.
initializer: `str` or `Tensor`. The variable initialization.
regularizer: `str` or `Tensor`. The variable regularizer.
trainable: `bool`. If True, this variable weights will be trained.
collections: `str`. A collection to add the new variable to (optional).
device: `str`. Device ID to store the variable. Default: '/cpu:0'.
restore: `bool`. Restore or not this variable when loading a pre-trained model.
Returns:
A Variable.
"""
if isinstance(initializer, six.string_types):
initializer = getters.get_initializer(initializer)
# Remove shape param if initializer is a Tensor
if not callable(initializer) and isinstance(initializer, tf.Tensor):
shape = None
if isinstance(regularizer, six.string_types):
regularizer = getters.get_regularizer(regularizer)
with tf.device(device_name_or_function=device):
var = tf.get_variable(name=name,
shape=shape,
dtype=dtype,
initializer=initializer,
regularizer=regularizer,
trainable=trainable,
collections=collections)
if not restore:
tf.add_to_collection(name=tf.GraphKeys.EXCL_RESTORE_VARIABLES, value=var) # @TODO adapt restoring saver
return var
