Rename objects to follow reference literature

cell.py

@@ -25,33 +25,33 @@ def state_size(self):
  def output_size(self):
  return self._size
 
- def __call__(self, input, state, scope=None):
+ def __call__(self, x, h, scope=None):
  with tf.variable_scope(scope or self.__class__.__name__):
- previous_memory, previous_output = state
+ previous_memory, previous_output = h
 
  with tf.variable_scope('Expand'):
- samples = input.get_shape()[0].value
+ samples = x.get_shape()[0].value
  shape = [samples, self._height, self._width, -1]
- input = tf.reshape(input, shape)
+ x = tf.reshape(x, shape)
  previous_memory = tf.reshape(previous_memory, shape)
  previous_output = tf.reshape(previous_output, shape)
 
  with tf.variable_scope('Convolve'):
- channels = input.get_shape()[-1].value
+ channels = x.get_shape()[-1].value
  filters = self._filters
  gates = 4 * filters if filters > 1 else 4
- x = tf.concat([input, previous_output], axis=3)
+ x = tf.concat([x, previous_output], axis=3)
  n = channels + filters
  m = gates
  W = tf.get_variable('kernel', self._kernel + [n, m], initializer=self._initializer)
  y = tf.nn.convolution(x, W, 'SAME')
  y += tf.get_variable('bias', [m], initializer=tf.constant_initializer(0.0))
- input, input_gate, forget_gate, output_gate = tf.split(y, 4, axis=3)
+ input_contribution, input_gate, forget_gate, output_gate = tf.split(y, 4, axis=3)
 
  with tf.variable_scope('LSTM'):
  memory = (previous_memory
  * tf.sigmoid(forget_gate + self._forget_bias)
- + tf.sigmoid(input_gate) * self._activation(input))
+ + tf.sigmoid(input_gate) * self._activation(input_contribution))
  output = self._activation(memory) * tf.sigmoid(output_gate)
 
  with tf.variable_scope('Flatten'):
@@ -82,32 +82,32 @@ def state_size(self):
  def output_size(self):
  return self._size
 
- def __call__(self, input, state, scope=None):
+ def __call__(self, x, h, scope=None):
  with tf.variable_scope(scope or self.__class__.__name__):
 
  with tf.variable_scope('Expand'):
- samples = input.get_shape()[0].value
+ samples = x.get_shape()[0].value
  shape = [samples, self._height, self._width, -1]
- input = tf.reshape(input, shape)
- state = tf.reshape(state, shape)
+ x = tf.reshape(x, shape)
+ h = tf.reshape(h, shape)
 
  with tf.variable_scope('Gates'):
  channels = input.get_shape()[-1].value
- x = tf.concat([input, state], axis=3)
+ inputs = tf.concat([x, h], axis=3)
  n = channels + self._filters
  m = 2 * self._filters if self._filters > 1 else 2
  W = tf.get_variable('kernel', self._kernel + [n, m], initializer=self._initializer)
- y = tf.nn.convolution(x, W, 'SAME')
+ y = tf.nn.convolution(inputs, W, 'SAME')
  y += tf.get_variable('bias', [m], initializer=tf.constant_initializer(1.0))
  reset_gate, update_gate = tf.split(y, 2, axis=3)
  reset_gate, update_gate = tf.sigmoid(reset_gate), tf.sigmoid(update_gate)
 
  with tf.variable_scope('Output'):
- x = tf.concat([input, reset_gate * state], axis=3)
+ inputs = tf.concat([x, reset_gate * h], axis=3)
  n = channels + self._filters
  m = self._filters
  W = tf.get_variable('kernel', self._kernel + [n, m], initializer=self._initializer)
- y = tf.nn.convolution(x, W, 'SAME')
+ y = tf.nn.convolution(inputs, W, 'SAME')
  y += tf.get_variable('bias', [m], initializer=tf.constant_initializer(0.0))
  y = self._activation(y)
  output = update_gate * state + (1 - update_gate) * y