def testTrainableFalseIsTrainingFalse(self):
embeddings = full_context_embeddings_cnn(self._images,
trainable=False,
is_training=False)
self.assertEqual([self._batch_size, 64],
embeddings.get_shape().as_list())
self._verifyParameterCounts()
self._assertCollectionSize(16, tf.GraphKeys.VARIABLES)
self._assertCollectionSize(0, tf.GraphKeys.TRAINABLE_VARIABLES)
self._assertCollectionSize(0, tf.GraphKeys.UPDATE_OPS)
self._assertCollectionSize(0, tf.GraphKeys.REGULARIZATION_LOSSES)
self._assertCollectionSize(0, tf.GraphKeys.LOSSES)
self._assertCollectionSize(4, tf.GraphKeys.SUMMARIES)
def build_fully_conditional_embedding_g(self):
"""Builds the fully conditional embedding g
Inputs:
self.sounds
Outputs:
self.g_embeddings
"""
model_output = full_context_embeddings_cnn(
self.support_set_sounds,
trainable=self.train_model,
is_training=self.is_training(),
scope="g_embedding_support_vectors")
self.g_model_variables = tf.get_collection(
tf.GraphKeys.VARIABLES, scope="g_embedding_support_vectors")
with tf.variable_scope("fce_embedding_g") as scope:
# sound_embeddings = tf.contrib.layers.fully_connected(
# inputs=model_output,
# num_outputs=self.config.embedding_size,
# activation_fn=None,
# weights_initializer=self.initializer,
# biases_initializer=None,
# scope=scope)
model_output = tf.expand_dims(model_output, 0)
model_output = tf.unstack(model_output)
cell_fw = tf.contrib.rnn.LSTMCell(self.config.embedding_size * 0.5,
initializer=self.initializer,
use_peepholes=True,
state_is_tuple=True)
# Backward direction cell
cell_bw = tf.contrib.rnn.LSTMCell(self.config.embedding_size * 0.5,
initializer=self.initializer,
use_peepholes=True,
state_is_tuple=True)
(outputs, state,
_) = tf.contrib.rnn.static_bidirectional_rnn(cell_fw,
cell_bw,
model_output,
dtype=tf.float32)
self.g_embedding = outputs
def build_fully_conditional_embedding_f(self):
"""Builds the fully conditional embedding f
Inputs:
self.sounds
Outputs:
self.sound_embeddings
"""
model_output = full_context_embeddings_cnn(
self.test_sound,
trainable=self.train_model,
is_training=self.is_training(),
scope="f_test_support_vector")
self.f_model_variables = tf.get_collection(
tf.GraphKeys.VARIABLES, scope="f_test_support_vector")
with tf.variable_scope('fce_embedding_f',
initializer=self.initializer) as scope:
# sound_embeddings = tf.contrib.layers.fully_connected(
# inputs=model_output,
# num_outputs=self.config.embedding_size,
# activation_fn=None,
# weights_initializer=self.initializer,
# biases_initializer=None,
# scope=scope)
# Feed the test image embeddings to set the initial LSTM state.
cell = tf.contrib.rnn.LSTMCell(
num_units=self.config.embedding_size,
state_is_tuple=False,
use_peepholes=True)
zero_state = cell.zero_state(
batch_size=model_output.get_shape()[0], dtype=tf.float32)
output, initial_state = cell(model_output, zero_state)
output = tf.add(model_output, output)
attention = tf.nn.softmax((tf.matmul(self.g_embedding[0],
tf.transpose(output))))
read_out = tf.reduce_sum(tf.matmul(tf.transpose(attention),
self.g_embedding[0]),
0,
keep_dims=True)
h_concatenated = tf.concat([output, read_out], 1)
scope.reuse_variables()
# Embedding shared by the input and outputs.
for i in xrange(self.config.lstm_processing_steps):
output, initial_state = cell(model_output, h_concatenated)
output = tf.add(model_output, output)
attention = tf.nn.softmax((tf.matmul(self.g_embedding[0],
tf.transpose(output))))
read_out = tf.reduce_sum(tf.matmul(tf.transpose(attention),
self.g_embedding[0]),
0,
keep_dims=True)
h_concatenated = tf.concat([output, read_out], 1)
self.f_embedding = output