def __call__(self, inputs, state, timestep=0, scope=None):
current_state = state
for highway_layer in xrange(self.num_highway_layers):
with tf.variable_scope('highway_factor_' + str(highway_layer)):
if self.use_inputs_on_each_layer or highway_layer == 0:
highway_factor = tf.tanh(
multiplicative_integration([inputs, current_state],
self._num_units))
else:
highway_factor = tf.tanh(
layer_norm(
linear([current_state], self._num_units, True)))
with tf.variable_scope('gate_for_highway_factor_' +
str(highway_layer)):
if self.use_inputs_on_each_layer or highway_layer == 0:
gate_for_highway_factor = tf.sigmoid(
multiplicative_integration([inputs, current_state],
self._num_units,
initial_bias_value=-3.0))
else:
gate_for_highway_factor = tf.sigmoid(
linear([current_state], self._num_units, True, -3.0))
gate_for_hidden_factor = 1 - gate_for_highway_factor
if self.use_recurrent_dropout and self.is_training:
highway_factor = tf.nn.dropout(
highway_factor, self.recurrent_dropout_factor)
current_state = highway_factor * gate_for_highway_factor + current_state * gate_for_hidden_factor
return current_state, current_state
def __call__(self, inputs, state, timestep=0, scope=None):
current_state = state
for highway_layer in xrange(self.num_highway_layers):
with tf.variable_scope('highway_factor_' + str(highway_layer)):
if self.use_inputs_on_each_layer or highway_layer == 0:
highway_factor = tf.tanh(
linear([inputs, current_state], self._num_units, True))
else:
highway_factor = tf.tanh(
linear([current_state], self._num_units, True))
with tf.variable_scope('gate_for_highway_factor_' +
str(highway_layer)):
if self.use_inputs_on_each_layer or highway_layer == 0:
gate_for_highway_factor = tf.sigmoid(
linear([inputs, current_state], self._num_units, True,
-3.0))
else:
gate_for_highway_factor = tf.sigmoid(
linear([current_state], self._num_units, True, -3.0))
gate_for_hidden_factor = 1.0 - gate_for_highway_factor
current_state = highway_factor * gate_for_highway_factor + current_state * gate_for_hidden_factor
return current_state, current_state
def __call__(self, inputs, state, timestep=0, scope=None):
with tf.device("/gpu:" + str(self._gpu_for_layer)):
"""Long short-term memory cell (LSTM)."""
with tf.variable_scope(scope
or type(self).__name__): # "BasicLSTMCell"
# Parameters of gates are concatenated into one multiply for efficiency.
h, c = tf.split(1, 2, state)
concat = multiplicative_integration([inputs, h],
self._num_units * 4, 0.0)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = tf.split(1, 4, concat)
if self.use_recurrent_dropout and self.is_training:
input_contribution = tf.nn.dropout(
tf.tanh(j), self.recurrent_dropout_factor)
else:
input_contribution = tf.tanh(j)
new_c = c * tf.sigmoid(f + self._forget_bias) + tf.sigmoid(
i) * input_contribution
new_h = tf.tanh(new_c) * tf.sigmoid(o)
return new_h, tf.concat(1, [new_h, new_c]) # purposely reversed
def __call__(self, inputs, state, timestep=0, scope=None):
with tf.variable_scope(scope
or type(self).__name__): # "BasicLSTMCell"
# Parameters of gates are concatenated into one multiply for efficiency.
hidden_state_plus_c_list = tf.split(1, self.num_memory_arrays + 1,
state)
h = hidden_state_plus_c_list[0]
c_list = hidden_state_plus_c_list[1:]
'''very large matrix multiplication to speed up procedure -- will split variables out later'''
if self.use_multiplicative_integration:
concat = multiplicative_integration(
[inputs, h], self._num_units * 4 * self.num_memory_arrays,
0.0)
else:
concat = linear([inputs, h],
self._num_units * 4 * self.num_memory_arrays,
True)
if self.use_layer_normalization:
concat = layer_norm(concat,
num_variables_in_tensor=4 *
self.num_memory_arrays)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate -- comes in sets of fours
all_vars_list = tf.split(1, 4 * self.num_memory_arrays, concat)
'''memory array loop'''
new_c_list, new_h_list = [], []
for array_counter in xrange(self.num_memory_arrays):
i = all_vars_list[0 + array_counter * 4]
j = all_vars_list[1 + array_counter * 4]
f = all_vars_list[2 + array_counter * 4]
o = all_vars_list[3 + array_counter * 4]
if self.use_recurrent_dropout and self.is_training:
input_contribution = tf.nn.dropout(
tf.tanh(j), self.recurrent_dropout_factor)
else:
input_contribution = tf.tanh(j)
new_c_list.append(c_list[array_counter] *
tf.sigmoid(f + self._forget_bias) +
tf.sigmoid(i) * input_contribution)
if self.use_layer_normalization:
new_c = layer_norm(new_c_list[-1])
else:
new_c = new_c_list[-1]
new_h_list.append(tf.tanh(new_c) * tf.sigmoid(o))
'''sum all new_h components -- could instead do a mean -- but investigate that later'''
new_h = tf.add_n(new_h_list)
return new_h, tf.concat(1, [new_h] + new_c_list) # purposely reversed
def __call__(self, inputs, state, timestep=0, scope=None):
"""Normal Gated recurrent unit (GRU) with nunits cells."""
with tf.variable_scope(scope or type(self).__name__): # "GRUCell"
with tf.variable_scope("Gates"): # Reset gate and update gate.
# We start with bias of 1.0 to not reset and not udpate.
r, u = tf.split(
1, 2,
tf.sigmoid(
multiplicative_integration([inputs, state],
self._num_units * 2, 1.0)))
with tf.variable_scope(
"Candidate"
): # you need a different one because you're doing a new linear
# notice they have the activation/non-linear step right here!
c = tf.tanh(
multiplicative_integration([inputs, state],
self._num_units, 0.0))
if self.use_recurrent_dropout and self.is_training:
input_contribution = tf.nn.dropout(
c, self.recurrent_dropout_factor)
else:
input_contribution = c
new_h = u * state + (1 - u) * input_contribution
return new_h, new_h
def __call__(self, inputs, state, timestep=0, scope=None):
with tf.variable_scope(scope or type(self).__name__):
with tf.variable_scope(
"Gates"
): # Forget Gate bias starts as 1.0 -- TODO: double check if this is correct
if self.use_multiplicative_integration:
gated_factor = multiplicative_integration(
[inputs, state], self._num_units,
self.forget_bias_initialization)
else:
gated_factor = linear([inputs, state], self._num_units,
True,
self.forget_bias_initialization)
gated_factor = tf.sigmoid(gated_factor)
with tf.variable_scope("Candidate"):
c = tf.tanh(linear([inputs], self._num_units, True, 0.0))
if self.use_recurrent_dropout and self.is_training:
input_contribution = tf.nn.dropout(
c, self.recurrent_dropout_factor)
else:
input_contribution = c
new_h = (1 -
gated_factor) * state + gated_factor * input_contribution
return new_h, new_h
def apply_highway_gate(proposed_output, original_input, bias=-2.0):
'''will apply a sigmoid transform_gate to any proposed output'''
transform_gate = tf.sigmoid(
linear_function(original_input,
proposed_output.get_shape()[1],
True,
bias,
scope='transform_lin_%d' % idx))
carry_gate = 1.0 - transform_gate
output = transform_gate * proposed_output + carry_gate * original_input
return output
def __call__(self, inputs, state, scope=None):
with tf.device("/gpu:" + str(self._gpu_for_layer)):
"""JZS1, mutant 1 with n units cells."""
with tf.variable_scope(scope or type(self).__name__): # "JZS1Cell"
with tf.variable_scope(
"Zinput"): # Reset gate and update gate.
# We start with bias of 1.0 to not reset and not update.
'''equation 1 z = sigm(WxzXt+Bz), x_t is inputs'''
z = tf.sigmoid(
linear([inputs],
self._num_units,
True,
1.0,
weight_initializer=self._weight_initializer,
orthogonal_scale_factor=self.
_orthogonal_scale_factor))
with tf.variable_scope("Rinput"):
'''equation 2 r = sigm(WxrXt+Whrht+Br), h_t is the previous state'''
r = tf.sigmoid(
linear([inputs, state],
self._num_units,
True,
1.0,
weight_initializer=self._weight_initializer,
orthogonal_scale_factor=self.
_orthogonal_scale_factor))
'''equation 3'''
with tf.variable_scope("Candidate"):
component_0 = linear([r * state], self._num_units, True)
component_1 = tf.tanh(tf.tanh(inputs) + component_0)
component_2 = component_1 * z
component_3 = state * (1 - z)
h_t = component_2 + component_3
return h_t, h_t # there is only one hidden state output to keep track of.
def highway(input_,
output_size,
num_layers=2,
bias=-2.0,
activation=tf.nn.relu,
scope=None,
use_batch_timesteps=False,
use_l2_loss=True,
timestep=-1):
"""Highway Network (cf. http://arxiv.org/abs/1505.00387).
t = sigmoid(Wy + b)
z = t * g(Wy + b) + (1 - t) * y
where g is nonlinearity, t is transform gate, and (1 - t) is carry gate.
if you initially set the bias to -2, then you achieve a simple pass through layer
use_batch_timesteps requires input to be 3d input [batch_size x timesteps x input_size] and will return a tensor of the exact same dimensions
"""
if output_size == 'same': output_size = input_.get_shape()[-1]
linear_function = linear.batch_timesteps_linear if use_batch_timesteps else linear.linear
with tf.variable_scope(scope or 'highway_network'):
output = input_
for idx in xrange(num_layers):
original_input = output
transform_gate = tf.sigmoid(
linear_function(original_input,
output_size,
True,
bias,
scope='transform_lin_%d' % idx,
timestep=timestep))
proposed_output = activation(
linear_function(original_input,
output_size,
True,
use_l2_loss=use_l2_loss,
scope='proposed_output_lin_%d' % idx,
timestep=timestep),
'activation_output_lin_' + str(idx))
carry_gate = 1.0 - transform_gate
output = transform_gate * proposed_output + carry_gate * original_input
return output
def __call__(self, inputs, state, scope=None):
with tf.device("/gpu:" + str(self._gpu_for_layer)):
"""JZS3, mutant 2 with n units cells."""
with tf.variable_scope(scope or type(self).__name__): # "JZS1Cell"
with tf.variable_scope(
"Zinput"): # Reset gate and update gate.
# We start with bias of 1.0 to not reset and not update.
'''equation 1'''
z = tf.sigmoid(
linear([inputs, tf.tanh(state)],
self._num_units,
True,
1.0,
weight_initializer=self._weight_initializer,
orthogonal_scale_factor=self.
_orthogonal_scale_factor))
'''equation 2'''
with tf.variable_scope("Rinput"):
r = tf.sigmoid(
linear([inputs, state],
self._num_units,
True,
1.0,
weight_initializer=self._weight_initializer,
orthogonal_scale_factor=self.
_orthogonal_scale_factor))
'''equation 3'''
with tf.variable_scope("Candidate"):
component_0 = linear([state * r, inputs], self._num_units,
True)
component_2 = (tf.tanh(component_0)) * z
component_3 = state * (1 - z)
h_t = component_2 + component_3
return h_t, h_t # there is only one hidden state output to keep track of.
def __call__(self, inputs, state, timestep=0, scope=None):
"""Long short-term memory cell (LSTM).
The idea with iteration would be to run different batch norm mean and variance stats on timestep greater than 10
"""
with tf.variable_scope(scope
or type(self).__name__): # "BasicLSTMCell"
# Parameters of gates are concatenated into one multiply for efficiency.
h, c = tf.split(1, 2, state)
'''note that bias is set to 0 because batch norm bias is added later'''
with tf.variable_scope('inputs_weight_matrix'):
inputs_concat = linear([inputs], 4 * self._num_units, False)
inputs_concat = layer_norm(inputs_concat,
num_variables_in_tensor=4,
scope="inputs_concat_layer_norm")
with tf.variable_scope('state_weight_matrix'):
h_concat = linear([h], 4 * self._num_units, False)
h_concat = layer_norm(h_concat,
num_variables_in_tensor=4,
scope="h_concat_layer_norm")
i, j, f, o = tf.split(
1, 4,
multiplicative_integration([inputs_concat, h_concat],
4 * self._num_units,
0.0,
weights_already_calculated=True))
new_c = c * tf.sigmoid(f + self._forget_bias) + tf.sigmoid(
i) * tf.tanh(j)
'''apply layer norm to the hidden state transition'''
with tf.variable_scope('layer_norm_hidden_state'):
new_h = tf.tanh(layer_norm(new_c)) * tf.sigmoid(o)
return new_h, tf.concat(1, [new_h, new_c]) # reversed this
def __call__(self, inputs, state, timestep=0, scope=None):
"""Normal Gated recurrent unit (GRU) with nunits cells."""
with tf.variable_scope(scope or type(self).__name__): # "GRUCell"
with tf.variable_scope("Inputs"):
inputs_concat = linear([inputs], self._num_units * 2, False,
1.0)
inputs_concat = layer_norm(inputs_concat,
num_variables_in_tensor=2,
initial_bias_value=1.0)
with tf.variable_scope("Hidden_State"):
hidden_state_concat = linear([state], self._num_units * 2,
False)
hidden_state_concat = layer_norm(hidden_state_concat,
num_variables_in_tensor=2)
r, u = tf.split(
1, 2,
tf.sigmoid(
multiplicative_integration(
[inputs_concat, hidden_state_concat],
2 * self._num_units,
1.0,
weights_already_calculated=True)))
with tf.variable_scope("Candidate"):
with tf.variable_scope('input_portion'):
input_portion = layer_norm(
linear([inputs], self._num_units, False))
with tf.variable_scope('reset_portion'):
reset_portion = r * layer_norm(
linear([state], self._num_units, False))
c = tf.tanh(
multiplicative_integration(
[input_portion, reset_portion],
self._num_units,
0.0,
weights_already_calculated=True))
new_h = u * state + (1 - u) * c
return new_h, new_h
def __call__(self, inputs, state, scope=None):
"""Gated recurrent unit (GRU) with nunits cells."""
with tf.variable_scope(scope or type(self).__name__): # "GRUCell"
with tf.variable_scope("Gates"): # Reset gate and update gate.
# We start with bias of 1.0 to not reset and not update.
concated_r_u = layer_norm(linear([inputs, state], 2 * self._num_units, False, 1.0),
num_variables_in_tensor=2, initial_bias_value=1.0)
r, u = tf.split(1, 2, tf.sigmoid(concated_r_u))
with tf.variable_scope("Candidate"):
with tf.variable_scope("reset_portion"):
reset_portion = r * layer_norm(linear([state], self._num_units, False))
with tf.variable_scope("inputs_portion"):
inputs_portion = layer_norm(linear([inputs], self._num_units, False))
c = tf.tanh(reset_portion + inputs_portion)
new_h = u * state + (1 - u) * c
return new_h, new_h
