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.
c, h = state
concat = linear([inputs, h], self._num_units * 4, False, 0.0)
concat = layers.layer_norm(concat, reuse=None)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = tf.split(concat, 4, 1)
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
with tf.variable_scope('new_h_output'):
new_h = tf.tanh(layers.layer_norm(
new_c, reuse=None)) * tf.sigmoid(o)
new_state = rnn_cell_impl.LSTMStateTuple(new_c, new_h)
# return new_h, tf.concat([new_h, new_c], 1) #purposely reversed
return new_h, new_state
def layer_norm(in_tensor, data_format='NHWC'):
if data_format == 'NCHW':
in_tensor = tf.transpose(in_tensor, (0, 2, 3, 1))
out_tensor = layer_norm(in_tensor)
if data_format == 'NCHW':
out_tensor = tf.transpose(out_tensor, (0, 3, 1, 2))
return out_tensor
def _norm(self, inp, scope):
with vs.variable_scope(scope) as scope:
shape = inp.get_shape()[-1:]
gamma_init = init_ops.constant_initializer(self._g)
beta_init = init_ops.constant_initializer(self._b)
gamma = vs.get_variable("gamma", shape=shape, initializer=gamma_init) # pylint: disable=unused-variable
beta = vs.get_variable("beta", shape=shape, initializer=beta_init) # pylint: disable=unused-variable
normalized = layers.layer_norm(inp, reuse=True, scope=scope)
return normalized
def _norm(self, inp, scope):
shape = inp.get_shape()[-1:]
gamma_init = init_ops.constant_initializer(self._g)
beta_init = init_ops.constant_initializer(self._b)
with vs.variable_scope(scope):
# Initialize beta and gamma for use by layer_norm.
vs.get_variable("gamma", shape=shape, initializer=gamma_init)
vs.get_variable("beta", shape=shape, initializer=beta_init)
normalized = layers.layer_norm(inp, reuse=True, scope=scope)
return normalized
def _norm(self, inp, scope):
shape = inp.get_shape()[-1:]
gamma_init = init_ops.constant_initializer(self._g)
beta_init = init_ops.constant_initializer(self._b)
with vs.variable_scope(scope):
# Initialize beta and gamma for use by layer_norm.
vs.get_variable("gamma", shape=shape, initializer=gamma_init)
vs.get_variable("beta", shape=shape, initializer=beta_init)
normalized = layers.layer_norm(inp, reuse=True, scope=scope)
return normalized
def layer(x, name=None):
with variable_scope.variable_scope(name, default_name="layer"):
x = layers.layer_norm(x)
x = convolutional.conv1d(
x,
10,
1,
use_bias=False,
kernel_initializer=init_ops.constant_initializer(42.42))
x = nn_ops.relu(x)
return x
def layer(x, name=None):
with variable_scope.variable_scope(name, default_name="layer"):
x = layers.layer_norm(x)
x = convolutional.conv1d(
x,
10,
1,
use_bias=False,
kernel_initializer=init_ops.constant_initializer(42.42))
x = nn_ops.relu(x)
return x
def _norm(self, inp, scope="layer_norm"):
""" TODO
"""
shape = inp.get_shape()[-1:]
gamma_init = init_ops.constant_initializer(1)
beta_init = init_ops.constant_initializer(1)
with tf.variable_scope(scope):
tf.get_variable("gamma", shape=shape, initializer=gamma_init)
tf.get_variable("beta", shape=shape, initializer=beta_init)
normalized = layers.layer_norm(inp, reuse=True, scope=scope)
return normalized
def _norm(self, inp, scope="layer_norm"):
"""
"""
shape = inp.get_shape()[-1:]
gamma_init = init_ops.constant_initializer(self._g)
beta_init = init_ops.constant_initializer(self._b)
with vs.variable_scope(scope, reuse=tf.AUTO_REUSE):
vs.get_variable("gamma", shape=shape, initializer=gamma_init)
vs.get_variable("beta", shape=shape, initializer=beta_init)
normalized = layers.layer_norm(inp, reuse=True, scope=scope)
return normalized
def _norm(self, inp, scope="layer_norm"):
"""
Performs layer normalization on the hidden state.
inp = the input to be normalized
Returns inp normalized by learned parameters gamma and beta
"""
shape = inp.get_shape()[-1:]
gamma_init = init_ops.constant_initializer(self._g)
beta_init = init_ops.constant_initializer(self._b)
with vs.variable_scope(scope):
vs.get_variable("gamma", shape=shape, initializer=gamma_init)
vs.get_variable("beta", shape=shape, initializer=beta_init)
normalized = layers.layer_norm(inp, reuse=True, scope=scope)
return normalized
def _norm(self, inp, scope="layer_norm"):
"""
Performs layer normalization on the hidden state.
inp = the input to be normalized
scope = name for the variable scope, just leave as default
Returns inp normalized by learned parameters gamma and beta
"""
shape = inp.get_shape()[-1:]
gamma_init = init_ops.constant_initializer(1)
beta_init = init_ops.constant_initializer(1)
with tf.variable_scope(scope):
tf.get_variable("gamma", shape=shape, initializer=gamma_init)
tf.get_variable("beta", shape=shape, initializer=beta_init)
normalized = layers.layer_norm(inp, reuse=True, scope=scope)
return normalized
def _norm(self, inp, scope):
return layers.layer_norm(inp, reuse=True, scope=scope)
