def __init__(self,
num_units,
input_size=None,
activation=tf.tanh,
inner_activation=tf.sigmoid,
bias=True,
weights_init=None,
trainable=True,
restore=True,
reuse=False):
if input_size is not None:
logging.warn("%s: The input_size parameter is deprecated." % self)
self._num_units = num_units
if isinstance(activation, str):
self._activation = activations.get(activation)
elif hasattr(activation, '__call__'):
self._activation = activation
else:
raise ValueError("Invalid Activation.")
if isinstance(inner_activation, str):
self._inner_activation = activations.get(inner_activation)
elif hasattr(inner_activation, '__call__'):
self._inner_activation = inner_activation
else:
raise ValueError("Invalid Activation.")
self.bias = bias
self.weights_init = weights_init
if isinstance(weights_init, str):
self.weights_init = initializations.get(weights_init)()
self.trainable = trainable
self.restore = restore
self.reuse = reuse
def activation(incoming, activation='linear', name='activation'):
""" Activation.
Apply given activation to incoming tensor.
Arguments:
incoming: A `Tensor`. The incoming tensor.
activation: `str` (name) or `function` (returning a `Tensor`).
Activation applied to this layer (see tflearn.activations).
Default: 'linear'.
"""
if isinstance(activation, str):
x = activations.get(activation)(incoming)
elif hasattr(incoming, '__call__'):
x = activation(incoming)
else:
raise ValueError('Unknown activation type.')
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, x)
return x
def activation(incoming, activation='linear', name='activation'):
""" Activation.
Apply given activation to incoming tensor.
Arguments:
incoming: A `Tensor`. The incoming tensor.
activation: `str` (name) or `function` (returning a `Tensor`).
Activation applied to this layer (see tflearn.activations).
Default: 'linear'.
"""
if isinstance(activation, str):
x = activations.get(activation)(incoming)
elif hasattr(activation, '__call__'):
x = activation(incoming)
else:
raise ValueError('Unknown activation type.')
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, x)
return x
def __init__(self,
num_units,
forget_bias=1.0,
input_size=None,
state_is_tuple=True,
activation=tf.tanh,
inner_activation=tf.sigmoid,
bias=True,
weights_init=None,
trainable=True,
restore=True,
reuse=False,
batch_norm=False):
if not state_is_tuple:
logging.warn(
"%s: Using a concatenated state is slower and will soon be "
"deprecated. Use state_is_tuple=True." % self)
if input_size is not None:
logging.warn("%s: The input_size parameter is deprecated." % self)
self._num_units = num_units
self._forget_bias = forget_bias
self._state_is_tuple = state_is_tuple
self.batch_norm = batch_norm
if isinstance(activation, str):
self._activation = activations.get(activation)
elif hasattr(activation, '__call__'):
self._activation = activation
else:
raise ValueError("Invalid Activation.")
if isinstance(inner_activation, str):
self._inner_activation = activations.get(inner_activation)
elif hasattr(inner_activation, '__call__'):
self._inner_activation = inner_activation
else:
raise ValueError("Invalid Activation.")
self.bias = bias
self.weights_init = weights_init
if isinstance(weights_init, str):
self.weights_init = initializations.get(weights_init)()
self.trainable = trainable
self.restore = restore
self.reuse = reuse
def block17(net, scale=1.0, activation="relu"):
tower_conv = relu(
batch_normalization(
conv_2d(net,
192,
1,
bias=False,
activation=None,
name='Conv2d_1x1')))
tower_conv_1_0 = relu(
batch_normalization(
conv_2d(net,
128,
1,
bias=False,
activation=None,
name='Conv2d_0a_1x1')))
tower_conv_1_1 = relu(
batch_normalization(
conv_2d(tower_conv_1_0,
160, [1, 7],
bias=False,
activation=None,
name='Conv2d_0b_1x7')))
tower_conv_1_2 = relu(
batch_normalization(
conv_2d(tower_conv_1_1,
192, [7, 1],
bias=False,
activation=None,
name='Conv2d_0c_7x1')))
tower_mixed = merge([tower_conv, tower_conv_1_2],
mode='concat',
axis=3)
tower_out = relu(
batch_normalization(
conv_2d(tower_mixed,
net.get_shape()[3],
1,
bias=False,
activation=None,
name='Conv2d_1x1')))
net += scale * tower_out
if activation:
if isinstance(activation, str):
net = activations.get(activation)(net)
elif hasattr(activation, '__call__'):
net = activation(net)
else:
raise ValueError("Invalid Activation.")
return net
def block8(net, scale=1.0, activation="relu"):
tower_conv = relu(batch_normalization(conv_2d(net, 192, 1, bias=False, activation=None, name='Conv2d_1x1')))
tower_conv1_0 = relu(batch_normalization(conv_2d(net, 192, 1, bias=False, activation=None, name='Conv2d_0a_1x1')))
tower_conv1_1 = relu(batch_normalization(conv_2d(tower_conv1_0, 224, [1,3], bias=False, activation=None, name='Conv2d_0b_1x3')))
tower_conv1_2 = relu(batch_normalization(conv_2d(tower_conv1_1, 256, [3,1], bias=False, name='Conv2d_0c_3x1')))
tower_mixed = merge([tower_conv,tower_conv1_2], mode='concat', axis=3)
tower_out = relu(batch_normalization(conv_2d(tower_mixed, net.get_shape()[3], 1, bias=False, activation=None, name='Conv2d_1x1')))
net += scale * tower_out
if activation:
if isinstance(activation, str):
net = activations.get(activation)(net)
elif hasattr(activation, '__call__'):
net = activation(net)
else:
raise ValueError("Invalid Activation.")
return net
def activation(incoming, activation='linear'):
""" Activation.
Apply given activation to incoming tensor.
Arguments:
incoming: A `Tensor`. The incoming tensor.
activation: `str` (name) or `Tensor`. Activation applied to this layer.
(see tflearn.activations). Default: 'linear'.
"""
if isinstance(activation, str):
return activations.get(activation)(incoming)
elif hasattr(incoming, '__call__'):
return activation(incoming)
else:
raise ValueError('Unknown activation type.')
def block8(net, scale=1.0, activation='relu'):
"""
"""
tower_conv = conv_2d(net, 192, 1, activation='relu',name='Conv2d_1x1')
tower_conv1_0 = conv_2d(net, 192, 1, activation='relu', name='Conv2d_0a_1x1')
tower_conv1_1 = conv_2d(tower_conv1_0, 224, [1,3], name='Conv2d_0b_1x3')
tower_conv1_2 = conv_2d(tower_conv1_1, 256, [3,1], name='Conv2d_0c_3x1')
tower_mixed = merge([tower_conv,tower_conv1_2], mode='concat', axis=3)
tower_out = conv_2d(tower_mixed, net.get_shape()[3], 1, activation=None, name='Conv2d_1x1')
net += scale * tower_out
if activation:
if isinstance(activation, str):
net = activations.get(activation)(net)
elif hasattr(activation, '__call__'):
net = activation(net)
else:
raise ValueError("Invalid Activation.")
return net
def activation(incoming, activation='linear'):
""" Activation.
Apply given activation to incoming tensor.
Arguments:
incoming: A `Tensor`. The incoming tensor.
activation: `str` (name) or `Tensor`. Activation applied to this layer.
(see tflearn.activations). Default: 'linear'.
"""
if isinstance(activation, str):
return activations.get(activation)(incoming)
elif hasattr(incoming, '__call__'):
return activation(incoming)
else:
raise ValueError('Unknown activation type.')
def highway(incoming, n_units, activation='linear', transform_dropout=None,
weights_init='truncated_normal', bias_init='zeros',
regularizer=None, weight_decay=0.001, trainable=True,
restore=True, reuse=False, scope=None,
name="FullyConnectedHighway"):
""" Fully Connected Highway.
A fully connected highway network layer, with some inspiration from
[https://github.com/fomorians/highway-fcn](https://github.com/fomorians/highway-fcn).
Input:
(2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Output:
2D Tensor [samples, n_units].
Arguments:
incoming: `Tensor`. Incoming (2+)D Tensor.
n_units: `int`, number of units for this layer.
activation: `str` (name) or `function` (returning a `Tensor`).
Activation applied to this layer (see tflearn.activations).
Default: 'linear'.
transform_dropout: `float`: Keep probability on the highway transform gate.
weights_init: `str` (name) or `Tensor`. Weights initialization.
(see tflearn.initializations) Default: 'truncated_normal'.
bias_init: `str` (name) or `Tensor`. Bias initialization.
(see tflearn.initializations) Default: 'zeros'.
regularizer: `str` (name) or `Tensor`. Add a regularizer to this
layer weights (see tflearn.regularizers). Default: None.
weight_decay: `float`. Regularizer decay parameter. Default: 0.001.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model
reuse: `bool`. If True and 'scope' is provided, this layer variables
will be reused (shared).
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: A name for this layer (optional). Default: 'FullyConnectedHighway'.
Attributes:
scope: `Scope`. This layer scope.
W: `Tensor`. Variable representing units weights.
W_t: `Tensor`. Variable representing units weights for transform gate.
b: `Tensor`. Variable representing biases.
b_t: `Tensor`. Variable representing biases for transform gate.
Links:
[https://arxiv.org/abs/1505.00387](https://arxiv.org/abs/1505.00387)
"""
input_shape = utils.get_incoming_shape(incoming)
assert len(input_shape) > 1, "Incoming Tensor shape must be at least 2-D"
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.variable_op_scope([incoming], scope, name, reuse=reuse) as scope:
name = scope.name
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = va.variable('W', shape=[n_inputs, n_units], regularizer=W_regul,
initializer=W_init, trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
if isinstance(bias_init, str):
bias_init = initializations.get(bias_init)()
b = va.variable('b', shape=[n_units], initializer=bias_init,
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
# Weight and bias for the transform gate
W_T = va.variable('W_T', shape=[n_inputs, n_units],
regularizer=None, initializer=W_init,
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W_T)
b_T = va.variable('b_T', shape=[n_units],
initializer=tf.constant_initializer(-1),
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b_T)
# If input is not 2d, flatten it.
if len(input_shape) > 2:
incoming = tf.reshape(incoming, [-1, n_inputs])
if isinstance(activation, str):
activation = activations.get(activation)
elif hasattr(activation, '__call__'):
activation = activation
else:
raise ValueError("Invalid Activation.")
H = activation(tf.matmul(incoming, W) + b)
T = tf.sigmoid(tf.matmul(incoming, W_T) + b_T)
if transform_dropout:
T = dropout(T, transform_dropout)
C = tf.sub(1.0, T)
inference = tf.add(tf.mul(H, T), tf.mul(incoming, C))
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.W_t = W_T
inference.b = b
inference.b_t = b_T
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def conv_2d(incoming,
nb_filter,
filter_size,
strides=1,
padding='same',
activation='linear',
bias=True,
weights_init='uniform_scaling',
bias_init='zeros',
regularizer=None,
weight_decay=0.001,
trainable=True,
restore=True,
reuse=False,
scope=None,
name="Conv2D"):
input_shape = utils.get_incoming_shape(incoming)
assert len(input_shape) == 4, "Incoming Tensor shape must be 4-D"
filter_size = utils.autoformat_filter_conv2d(filter_size, input_shape[-1],
nb_filter)
strides = utils.autoformat_kernel_2d(strides)
padding = utils.autoformat_padding(padding)
with tf.variable_scope(scope,
default_name=name,
values=[incoming],
reuse=reuse) as scope:
name = scope.name
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer is not None:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = vs.variable('W',
shape=filter_size,
regularizer=W_regul,
initializer=W_init,
trainable=trainable,
restore=restore)
# Track per layer variables
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
b = None
if bias:
if isinstance(bias_init, str):
bias_init = initializations.get(bias_init)()
b = vs.variable('b',
shape=nb_filter,
initializer=bias_init,
trainable=trainable,
restore=restore)
# Track per layer variables
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
inference = tf.nn.conv2d(incoming, W, strides, padding)
if b is not None: inference = tf.nn.bias_add(inference, b)
if activation:
if isinstance(activation, str):
inference = activations.get(activation)(inference)
elif hasattr(activation, '__call__'):
inference = activation(inference)
else:
raise ValueError("Invalid Activation.")
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def single_unit(incoming, activation='linear', bias=True, trainable=True,
restore=True, reuse=False, scope=None, name="Linear"):
""" Single Unit.
A single unit (Linear) Layer.
Input:
1-D Tensor [samples]. If not 2D, input will be flatten.
Output:
1-D Tensor [samples].
Arguments:
incoming: `Tensor`. Incoming Tensor.
activation: `str` (name) or `function`. Activation applied to this
layer (see tflearn.activations). Default: 'linear'.
bias: `bool`. If True, a bias is used.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model.
reuse: `bool`. If True and 'scope' is provided, this layer variables
will be reused (shared).
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: A name for this layer (optional). Default: 'Linear'.
Attributes:
W: `Tensor`. Variable representing weight.
b: `Tensor`. Variable representing bias.
"""
input_shape = utils.get_incoming_shape(incoming)
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.variable_op_scope([incoming], scope, name, reuse=reuse) as scope:
name = scope.name
W = va.variable('W', shape=[n_inputs],
initializer=tf.constant_initializer(np.random.randn()),
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
b = None
if bias:
b = va.variable('b', shape=[n_inputs],
initializer=tf.constant_initializer(np.random.randn()),
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 1:
inference = tf.reshape(inference, [-1])
inference = tf.mul(inference, W)
if b: inference = tf.add(inference, b)
if isinstance(activation, str):
inference = activations.get(activation)(inference)
elif hasattr(activation, '__call__'):
inference = activation(inference)
else:
raise ValueError("Invalid Activation.")
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def fully_connected(incoming,
n_units,
activation='linear',
bias=True,
weights_init='truncated_normal',
bias_init='zeros',
regularizer=None,
weight_decay=0.001,
trainable=True,
restore=True,
name="FullyConnected"):
""" Fully Connected.
A fully connected layer.
Input:
(2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Output:
2D Tensor [samples, n_units].
Arguments:
incoming: `Tensor`. Incoming (2+)D Tensor.
n_units: `int`, number of units for this layer.
activation: `str` (name) or `Tensor`. Activation applied to this layer.
(see tflearn.activations). Default: 'linear'.
bias: `bool`. If True, a bias is used.
weights_init: `str` (name) or `Tensor`. Weights initialization.
(see tflearn.initializations) Default: 'truncated_normal'.
bias_init: `str` (name) or `Tensor`. Bias initialization.
(see tflearn.initializations) Default: 'zeros'.
regularizer: `str` (name) or `Tensor`. Add a regularizer to this
layer weights (see tflearn.regularizers). Default: None.
weight_decay: `float`. Regularizer decay parameter. Default: 0.001.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model
name: A name for this layer (optional). Default: 'FullyConnected'.
Attributes:
scope: `Scope`. This layer scope.
W: `Tensor`. Variable representing units weights.
b: `Tensor`. Variable representing biases.
"""
input_shape = utils.get_incoming_shape(incoming)
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.name_scope(name) as scope:
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = va.variable(scope + 'W',
shape=[n_inputs, n_units],
regularizer=W_regul,
initializer=W_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, W)
b = None
if bias:
b_init = initializations.get(bias_init)()
b = va.variable(scope + 'b',
shape=[n_units],
initializer=b_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 2:
inference = tf.reshape(inference, [-1, n_inputs])
inference = tf.matmul(inference, W)
if b: inference = tf.nn.bias_add(inference, b)
inference = activations.get(activation)(inference)
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
return inference
def fully_connected(incoming,
n_units,
activation='linear',
bias=True,
weights_init='truncated_normal',
bias_init='zeros',
regularizer=None,
weight_decay=0.001,
trainable=True,
restore=True,
reuse=False,
scope=None,
name="FullyConnected"):
input_shape = utils.get_incoming_shape(incoming)
assert len(input_shape) > 1, "Incoming Tensor shape must be at least 2-D"
n_inputs = int(np.prod(input_shape[1:]))
with tf.variable_scope(scope,
default_name=name,
values=[incoming],
reuse=reuse) as scope:
name = scope.name
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer is not None:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = vs.variable('W',
shape=[n_inputs, n_units],
regularizer=W_regul,
initializer=W_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
b = None
if bias:
if isinstance(bias_init, str):
bias_init = initializations.get(bias_init)()
b = vs.variable('b',
shape=[n_units],
initializer=bias_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 2:
inference = tf.reshape(inference, [-1, n_inputs])
inference = tf.matmul(inference, W)
if b is not None: inference = tf.nn.bias_add(inference, b)
if activation:
if isinstance(activation, str):
inference = activations.get(activation)(inference)
elif hasattr(activation, '__call__'):
inference = activation(inference)
else:
raise ValueError("Invalid Activation.")
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def single_unit(incoming,
activation='linear',
bias=True,
trainable=True,
restore=True,
name="Linear"):
""" Single Unit.
A single unit (Linear) Layer.
Input:
1-D Tensor [samples]. If not 2D, input will be flatten.
Output:
1-D Tensor [samples].
Arguments:
incoming: `Tensor`. Incoming Tensor.
activation: `str` (name) or `Tensor`. Activation applied to this layer.
(see tflearn.activations). Default: 'linear'.
bias: `bool`. If True, a bias is used.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model.
name: A name for this layer (optional). Default: 'Dense'.
Attributes:
W: `Tensor`. Variable representing weight.
b: `Tensor`. Variable representing bias.
"""
input_shape = utils.get_incoming_shape(incoming)
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.name_scope(name) as scope:
W = va.variable(scope + 'W',
shape=[n_inputs],
initializer=tf.constant_initializer(np.random.randn()),
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, W)
b = None
if bias:
b = va.variable(scope + 'b',
shape=[n_inputs],
initializer=tf.constant_initializer(
np.random.randn()),
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 1:
inference = tf.reshape(inference, [-1])
inference = tf.mul(inference, W)
if b: inference = tf.add(inference, b)
inference = activations.get(activation)(inference)
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
return inference
def PhraseLayer(incoming,
input_dim,
output_dim,
output_length,
activation='linear',
dropout_keepprob=0.5,
batchNorm=False,
name='PhraseLayer',
alpha=0.5,
scope=None):
'''
incoming: [batch_size, sen_length, input_dim]
return: [batch_size, sen_length, output_length, output_dim[0]],
[batch_size, sen_length, output_length, output_dim[1]]
'''
with tf.variable_scope(scope, default_name=name,
values=[incoming]) as scope:
name = scope.name
P = va.variable('P',
shape=[input_dim, output_dim[0]],
initializer=initializations.get('truncated_normal')())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, P)
P_p = va.variable(
'P_p',
shape=[input_dim, output_dim[1]],
initializer=initializations.get('truncated_normal')())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, P_p)
Q = va.variable('Q',
shape=[input_dim, output_dim[0]],
initializer=initializations.get('truncated_normal')())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, Q)
Q_p = va.variable(
'Q_p',
shape=[input_dim, output_dim[1]],
initializer=initializations.get('truncated_normal')())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, Q_p)
R = va.variable('R',
shape=[input_dim, output_dim[0]],
initializer=initializations.get('truncated_normal')())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, R)
R_p = va.variable(
'R_p',
shape=[input_dim, output_dim[1]],
initializer=initializations.get('truncated_normal')())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, R_p)
O = va.variable('O',
shape=[output_dim[0], output_dim[0]],
initializer=initializations.get('truncated_normal')())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, O)
O_p = va.variable('O_p',
shape=[output_dim[1], output_dim[1]],
initializer=tf.ones_initializer())
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, O_p)
b = va.variable('b',
shape=[1, 1, 1, output_dim[0]],
initializer=tf.ones_initializer())
b_p = va.variable('b_p',
shape=[1, 1, 1, output_dim[1]],
initializer=tf.ones_initializer())
if isinstance(activation, str):
activation = activations.get(activation)
elif hasattr(activation, '__call__'):
activation = activation
else:
raise ValueError("Invalid Activation.")
def calc(incoming, P, Q, R, O, output_dim):
batch_size = tf.shape(incoming)[0]
sent_length = incoming.shape[1].value
G1 = tf.zeros([batch_size, sent_length, output_dim])
G2 = tf.zeros([batch_size, sent_length, output_dim])
G3 = tf.zeros([batch_size, sent_length, output_dim])
r = []
for i in range(output_length):
'''if i == 0:
now = incoming
else:
now = tf.concat([tf.zeros([batch_size, i, input_dim]), incoming[:,0:-i, :]], axis = 1)
F2 = tf.einsum('aij,jk->aik', now, Q) * G1
F3 = tf.einsum('aij,jk->aik', now, R) * G2
if i == 0:
F1 = tf.einsum('aij,jk->aik', now, P)
G1 = G1 * alpha + F1
else:
G1 = G1 * alpha
G2 = G2 * alpha + F2
G3 = G3 * alpha + F3
r.append(tf.einsum('aij,jk->aik',G1+G2+G3, O))'''
F1 = tf.einsum('aij,jk->aik', incoming, P)
r.append(tf.einsum('aij,jk->aik', F1, O))
#return tf.stack(r, axis = 2)
return tf.reshape(r[0], [batch_size, sent_length, 1, output_dim])
batch_size = tf.shape(incoming)[0]
sent_length = incoming.shape[1].value
#out1 = tf.reshape(tf.einsum('aij,jk->aik', tf.einsum('aij,jk->aik', incoming, P), O), [batch_size, sent_length, 1, output_dim[0]]) + b
out1 = tf.reshape(tf.einsum('aij,jk->aik', incoming, P),
[batch_size, sent_length, 1, output_dim[0]]) + b
#out1 = calc(incoming, P, Q, R, O, output_dim[0]) + b
#out1 = activation(out1, name="activation")
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, out1)
if batchNorm:
pass
#out1 = tflearn.batch_normalization(out1, name="batchNormOut1")
#out1 = tflearn.dropout(out1, dropout_keepprob, name="dropOut1")
if output_dim[1] == 0:
out2 = None
else:
out2 = calc(tf.stop_gradient(incoming), P_p, Q_p, R_p, O_p,
output_dim[1]) + b_p
out2 = activation(out2, name="activation_p")
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, out2)
if batchNorm:
out2 = tflearn.batch_normalization(out2, name="batchNormOut2")
out2 = tflearn.dropout(out2, dropout_keepprob, name="dropOut2")
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, out1)
if output_dim[1] != 0:
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, out2)
return out1, out2
def conv_2d_BN(incoming, nb_filter, filter_size, strides=1, padding='same',
activation='linear', bias=True, weights_init='xavier',
bias_init='zeros', regularizer=None, weight_decay=0.001,
trainable=True, restore=True, reuse=False, scope=None,
name="Conv2D", batch_norm=False):
""" Convolution 2D.
Input:
4-D Tensor [batch, height, width, in_channels].
Output:
4-D Tensor [batch, new height, new width, nb_filter].
Arguments:
incoming: `Tensor`. Incoming 4-D Tensor.
nb_filter: `int`. The number of convolutional filters.
filter_size: `int` or `list of int`. Size of filters.
strides: 'int` or list of `int`. Strides of conv operation.
Default: [1 1 1 1].
padding: `str` from `"same", "valid"`. Padding algo to use.
Default: 'same'.
activation: `str` (name) or `function` (returning a `Tensor`).
Activation applied to this layer (see tflearn.activations).
Default: 'linear'.
bias: `bool`. If True, a bias is used.
weights_init: `str` (name) or `Tensor`. Weights initialization.
(see tflearn.initializations) Default: 'truncated_normal'.
bias_init: `str` (name) or `Tensor`. Bias initialization.
(see tflearn.initializations) Default: 'zeros'.
regularizer: `str` (name) or `Tensor`. Add a regularizer to this
layer weights (see tflearn.regularizers). Default: None.
weight_decay: `float`. Regularizer decay parameter. Default: 0.001.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model.
reuse: `bool`. If True and 'scope' is provided, this layer variables
will be reused (shared).
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: A name for this layer (optional). Default: 'Conv2D'.
batch_norm: If true, add batch normalization with default TFLearn
parameters before the activation layer
Attributes:
scope: `Scope`. This layer scope.
W: `Variable`. Variable representing filter weights.
b: `Variable`. Variable representing biases.
"""
input_shape = utils.get_incoming_shape(incoming)
assert len(input_shape) == 4, "Incoming Tensor shape must be 4-D"
filter_size = utils.autoformat_filter_conv2d(filter_size,
input_shape[-1],
nb_filter)
strides = utils.autoformat_kernel_2d(strides)
padding = utils.autoformat_padding(padding)
# Variable Scope fix for older TF
try:
vscope = tf.variable_scope(scope, default_name=name, values=[incoming],
reuse=reuse)
except Exception:
vscope = tf.variable_op_scope([incoming], scope, name, reuse=reuse)
with vscope as scope:
name = scope.name
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = vs.variable('W', shape=filter_size, regularizer=W_regul,
initializer=W_init, trainable=trainable,
restore=restore)
# Track per layer variables
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
b = None
if bias:
if isinstance(bias_init, str):
bias_init = initializations.get(bias_init)()
b = vs.variable('b', shape=nb_filter, initializer=bias_init,
trainable=trainable, restore=restore)
# Track per layer variables
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
inference = tf.nn.conv2d(incoming, W, strides, padding)
if b: inference = tf.nn.bias_add(inference, b)
if batch_norm:
inference = batch_normalization(inference)
if isinstance(activation, str):
if activation == 'softmax':
shapes = inference.get_shape()
inference = activations.get(activation)(inference)
elif hasattr(activation, '__call__'):
inference = activation(inference)
else:
raise ValueError("Invalid Activation.")
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def single_unit(incoming,
activation='linear',
bias=True,
trainable=True,
restore=True,
reuse=False,
scope=None,
name="Linear"):
""" Single Unit.
A single unit (Linear) Layer.
Input:
1-D Tensor [samples]. If not 2D, input will be flatten.
Output:
1-D Tensor [samples].
Arguments:
incoming: `Tensor`. Incoming Tensor.
activation: `str` (name) or `function`. Activation applied to this
layer (see tflearn.activations). Default: 'linear'.
bias: `bool`. If True, a bias is used.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model.
reuse: `bool`. If True and 'scope' is provided, this layer variables
will be reused (shared).
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: A name for this layer (optional). Default: 'Linear'.
Attributes:
W: `Tensor`. Variable representing weight.
b: `Tensor`. Variable representing bias.
"""
input_shape = utils.get_incoming_shape(incoming)
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.variable_scope(scope,
default_name=name,
values=[incoming],
reuse=reuse) as scope:
name = scope.name
W = va.variable('W',
shape=[n_inputs],
initializer=tf.constant_initializer(np.random.randn()),
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
b = None
if bias:
b = va.variable('b',
shape=[n_inputs],
initializer=tf.constant_initializer(
np.random.randn()),
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 1:
inference = tf.reshape(inference, [-1])
inference = tf.multiply(inference, W)
if b: inference = tf.add(inference, b)
if isinstance(activation, str):
inference = activations.get(activation)(inference)
elif hasattr(activation, '__call__'):
inference = activation(inference)
else:
raise ValueError("Invalid Activation.")
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def fully_connected(incoming, n_units, activation='linear', bias=True,
weights_init='truncated_normal', bias_init='zeros',
regularizer=None, weight_decay=0.001, trainable=True,
restore=True, reuse=False, scope=None,
name="FullyConnected"):
""" Fully Connected.
A fully connected layer.
Input:
(2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Output:
2D Tensor [samples, n_units].
Arguments:
incoming: `Tensor`. Incoming (2+)D Tensor.
n_units: `int`, number of units for this layer.
activation: `str` (name) or `function` (returning a `Tensor`).
Activation applied to this layer (see tflearn.activations).
Default: 'linear'.
bias: `bool`. If True, a bias is used.
weights_init: `str` (name) or `Tensor`. Weights initialization.
(see tflearn.initializations) Default: 'truncated_normal'.
bias_init: `str` (name) or `Tensor`. Bias initialization.
(see tflearn.initializations) Default: 'zeros'.
regularizer: `str` (name) or `Tensor`. Add a regularizer to this
layer weights (see tflearn.regularizers). Default: None.
weight_decay: `float`. Regularizer decay parameter. Default: 0.001.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model.
reuse: `bool`. If True and 'scope' is provided, this layer variables
will be reused (shared).
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: A name for this layer (optional). Default: 'FullyConnected'.
Attributes:
scope: `Scope`. This layer scope.
W: `Tensor`. Variable representing units weights.
b: `Tensor`. Variable representing biases.
"""
input_shape = utils.get_incoming_shape(incoming)
assert len(input_shape) > 1, "Incoming Tensor shape must be at least 2-D"
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.variable_op_scope([incoming], scope, name, reuse=reuse) as scope:
name = scope.name
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = va.variable('W', shape=[n_inputs, n_units], regularizer=W_regul,
initializer=W_init, trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
b = None
if bias:
if isinstance(bias, str):
bias_init = initializations.get(bias_init)()
b = va.variable('b', shape=[n_units], initializer=bias_init,
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 2:
inference = tf.reshape(inference, [-1, n_inputs])
inference = tf.matmul(inference, W)
if b: inference = tf.nn.bias_add(inference, b)
if isinstance(activation, str):
inference = activations.get(activation)(inference)
elif hasattr(activation, '__call__'):
inference = activation(inference)
else:
raise ValueError("Invalid Activation.")
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def fully_connected(incoming,
n_units,
activation='linear',
bias=True,
weights_init='truncated_normal',
bias_init='zeros',
regularizer=None,
weight_decay=0.001,
trainable=True,
restore=True,
reuse=False,
scope=None,
name="FullyConnected"):
""" Fully Connected.
A fully connected layer.
Input:
(2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Output:
2D Tensor [samples, n_units].
Arguments:
incoming: `Tensor`. Incoming (2+)D Tensor.
n_units: `int`, number of units for this layer.
activation: `str` (name) or `function` (returning a `Tensor`).
Activation applied to this layer (see tflearn.activations).
Default: 'linear'.
bias: `bool`. If True, a bias is used.
weights_init: `str` (name) or `Tensor`. Weights initialization.
(see tflearn.initializations) Default: 'truncated_normal'.
bias_init: `str` (name) or `Tensor`. Bias initialization.
(see tflearn.initializations) Default: 'zeros'.
regularizer: `str` (name) or `Tensor`. Add a regularizer to this
layer weights (see tflearn.regularizers). Default: None.
weight_decay: `float`. Regularizer decay parameter. Default: 0.001.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model.
reuse: `bool`. If True and 'scope' is provided, this layer variables
will be reused (shared).
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: A name for this layer (optional). Default: 'FullyConnected'.
Attributes:
scope: `Scope`. This layer scope.
W: `Tensor`. Variable representing units weights.
b: `Tensor`. Variable representing biases.
"""
input_shape = utils.get_incoming_shape(incoming)
assert len(input_shape) > 1, "Incoming Tensor shape must be at least 2-D"
n_inputs = int(np.prod(input_shape[1:]))
with tf.variable_scope(scope,
default_name=name,
values=[incoming],
reuse=reuse) as scope:
name = scope.name
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = va.variable('W',
shape=[n_inputs, n_units],
regularizer=W_regul,
initializer=W_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
b = None
if bias:
if isinstance(bias_init, str):
bias_init = initializations.get(bias_init)()
b = va.variable('b',
shape=[n_units],
initializer=bias_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 2:
inference = tf.reshape(inference, [-1, n_inputs])
inference = tf.matmul(inference, W)
if b: inference = tf.nn.bias_add(inference, b)
if activation:
if isinstance(activation, str):
inference = activations.get(activation)(inference)
elif hasattr(activation, '__call__'):
inference = activation(inference)
else:
raise ValueError("Invalid Activation.")
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def single_unit(incoming, activation='linear', bias=True, trainable=True,
restore=True, name="Linear"):
""" Single Unit.
A single unit (Linear) Layer.
Input:
1-D Tensor [samples]. If not 2D, input will be flatten.
Output:
1-D Tensor [samples].
Arguments:
incoming: `Tensor`. Incoming Tensor.
activation: `str` (name) or `Tensor`. Activation applied to this layer.
(see tflearn.activations). Default: 'linear'.
bias: `bool`. If True, a bias is used.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model.
name: A name for this layer (optional). Default: 'Dense'.
Attributes:
W: `Tensor`. Variable representing weight.
b: `Tensor`. Variable representing bias.
"""
input_shape = utils.get_incoming_shape(incoming)
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.name_scope(name) as scope:
W = va.variable(scope + 'W', shape=[n_inputs],
initializer=tf.constant_initializer(np.random.randn()),
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, W)
b = None
if bias:
b = va.variable(scope + 'b', shape=[n_inputs],
initializer=tf.constant_initializer(np.random.randn()),
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 1:
inference = tf.reshape(inference, [-1])
inference = tf.mul(inference, W)
if b: inference = tf.add(inference, b)
inference = activations.get(activation)(inference)
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
return inference
def highway(incoming,
n_units,
activation='linear',
transform_dropout=None,
weights_init='truncated_normal',
bias_init='zeros',
regularizer=None,
weight_decay=0.001,
trainable=True,
restore=True,
reuse=False,
scope=None,
name="FullyConnectedHighway"):
""" Fully Connected Highway.
A fully connected highway network layer, with some inspiration from
[https://github.com/fomorians/highway-fcn](https://github.com/fomorians/highway-fcn).
Input:
(2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Output:
2D Tensor [samples, n_units].
Arguments:
incoming: `Tensor`. Incoming (2+)D Tensor.
n_units: `int`, number of units for this layer.
activation: `str` (name) or `function` (returning a `Tensor`).
Activation applied to this layer (see tflearn.activations).
Default: 'linear'.
transform_dropout: `float`: Keep probability on the highway transform gate.
weights_init: `str` (name) or `Tensor`. Weights initialization.
(see tflearn.initializations) Default: 'truncated_normal'.
bias_init: `str` (name) or `Tensor`. Bias initialization.
(see tflearn.initializations) Default: 'zeros'.
regularizer: `str` (name) or `Tensor`. Add a regularizer to this
layer weights (see tflearn.regularizers). Default: None.
weight_decay: `float`. Regularizer decay parameter. Default: 0.001.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model
reuse: `bool`. If True and 'scope' is provided, this layer variables
will be reused (shared).
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: A name for this layer (optional). Default: 'FullyConnectedHighway'.
Attributes:
scope: `Scope`. This layer scope.
W: `Tensor`. Variable representing units weights.
W_t: `Tensor`. Variable representing units weights for transform gate.
b: `Tensor`. Variable representing biases.
b_t: `Tensor`. Variable representing biases for transform gate.
Links:
[https://arxiv.org/abs/1505.00387](https://arxiv.org/abs/1505.00387)
"""
input_shape = utils.get_incoming_shape(incoming)
assert len(input_shape) > 1, "Incoming Tensor shape must be at least 2-D"
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.variable_scope(scope,
default_name=name,
values=[incoming],
reuse=reuse) as scope:
name = scope.name
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = va.variable('W',
shape=[n_inputs, n_units],
regularizer=W_regul,
initializer=W_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W)
if isinstance(bias_init, str):
bias_init = initializations.get(bias_init)()
b = va.variable('b',
shape=[n_units],
initializer=bias_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b)
# Weight and bias for the transform gate
W_T = va.variable('W_T',
shape=[n_inputs, n_units],
regularizer=None,
initializer=W_init,
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, W_T)
b_T = va.variable('b_T',
shape=[n_units],
initializer=tf.constant_initializer(-1),
trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + name, b_T)
# If input is not 2d, flatten it.
if len(input_shape) > 2:
incoming = tf.reshape(incoming, [-1, n_inputs])
if isinstance(activation, str):
activation = activations.get(activation)
elif hasattr(activation, '__call__'):
activation = activation
else:
raise ValueError("Invalid Activation.")
H = activation(tf.matmul(incoming, W) + b)
T = tf.sigmoid(tf.matmul(incoming, W_T) + b_T)
if transform_dropout:
T = dropout(T, transform_dropout)
C = tf.subtract(1.0, T)
inference = tf.add(tf.multiply(H, T), tf.multiply(incoming, C))
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.W_t = W_T
inference.b = b
inference.b_t = b_T
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, inference)
return inference
def fully_connected(incoming, n_units, activation='linear', bias=True,
weights_init='truncated_normal', bias_init='zeros',
regularizer=None, weight_decay=0.001, trainable=True,
restore=True, name="FullyConnected"):
""" Fully Connected.
A fully connected layer.
Input:
(2+)-D Tensor [samples, input dim]. If not 2D, input will be flatten.
Output:
2D Tensor [samples, n_units].
Arguments:
incoming: `Tensor`. Incoming (2+)D Tensor.
n_units: `int`, number of units for this layer.
activation: `str` (name) or `Tensor`. Activation applied to this layer.
(see tflearn.activations). Default: 'linear'.
bias: `bool`. If True, a bias is used.
weights_init: `str` (name) or `Tensor`. Weights initialization.
(see tflearn.initializations) Default: 'truncated_normal'.
bias_init: `str` (name) or `Tensor`. Bias initialization.
(see tflearn.initializations) Default: 'zeros'.
regularizer: `str` (name) or `Tensor`. Add a regularizer to this
layer weights (see tflearn.regularizers). Default: None.
weight_decay: `float`. Regularizer decay parameter. Default: 0.001.
trainable: `bool`. If True, weights will be trainable.
restore: `bool`. If True, this layer weights will be restored when
loading a model
name: A name for this layer (optional). Default: 'FullyConnected'.
Attributes:
scope: `Scope`. This layer scope.
W: `Tensor`. Variable representing units weights.
b: `Tensor`. Variable representing biases.
"""
input_shape = utils.get_incoming_shape(incoming)
n_inputs = int(np.prod(input_shape[1:]))
# Build variables and inference.
with tf.name_scope(name) as scope:
W_init = weights_init
if isinstance(weights_init, str):
W_init = initializations.get(weights_init)()
W_regul = None
if regularizer:
W_regul = lambda x: losses.get(regularizer)(x, weight_decay)
W = va.variable(scope + 'W', shape=[n_inputs, n_units],
regularizer=W_regul, initializer=W_init,
trainable=trainable, restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, W)
b = None
if bias:
b_init = initializations.get(bias_init)()
b = va.variable(scope + 'b', shape=[n_units],
initializer=b_init, trainable=trainable,
restore=restore)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + scope, b)
inference = incoming
# If input is not 2d, flatten it.
if len(input_shape) > 2:
inference = tf.reshape(inference, [-1, n_inputs])
inference = tf.matmul(inference, W)
if b: inference = tf.nn.bias_add(inference, b)
inference = activations.get(activation)(inference)
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, inference)
# Add attributes to Tensor to easy access weights.
inference.scope = scope
inference.W = W
inference.b = b
return inference
