def _set_network(self):
"""Set up network.
Args:
void
Returns:
input (): Placeholder for network input.
output (): Output operation.
"""
# input
input = tf.placeholder(shape=self.s_dim, dtype=tf.float32)
# initializers for fc layers
unif_init = unif_initializer(-0.003, 0.003)
# flatten
net = tf.contrib.layers.flatten(input)
# fc layers
net = tf.layers.dense(net, 400, kernel_initializer=xav_init())
net = tf.nn.relu(net)
net = tf.layers.dense(net, 300, kernel_initializer=xav_init())
net = tf.nn.relu(net)
net = tf.layers.dense(net, self.a_dim[1], kernel_initializer=unif_init)
output = tf.nn.tanh(net)
return input, output
def _set_network(self):
"""Set up network.
Args:
void
Returns:
input (): Placeholder for network input.
actions (): Placeholder for action inputs.
output (): Output operation.
"""
# regularizer
regularizer = l2_regularizer(scale=0.01)
# inputs
input = tf.placeholder(shape=self.s_dim, dtype=tf.float32)
actions = tf.placeholder(shape=self.a_dim, dtype=tf.float32)
# initializers for fc layers
unif_init = unif_initializer(-0.003, 0.003)
# flatten
net = tf.contrib.layers.flatten(input)
# fc layers
net = tf.layers.dense(net,
400,
kernel_initializer=xav_init(),
kernel_regularizer=regularizer)
net = tf.nn.relu(net)
net = tf.layers.dense(net,
300,
kernel_initializer=xav_init(),
kernel_regularizer=regularizer,
use_bias=False)
action_net = tf.layers.dense(actions,
300,
kernel_initializer=xav_init(),
kernel_regularizer=regularizer,
use_bias=True)
net += action_net
net = tf.nn.relu(net)
output = tf.layers.dense(net,
1,
kernel_initializer=unif_init,
kernel_regularizer=regularizer)
return input, actions, output
def _vgg16(inputs, training=True, embedding_size=64,
dropout_keep_prob=0.5,
middleRepr=False,
scope='vgg_16'):
""" From https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/slim/python/slim/nets/vgg.py
Oxford Net VGG 16-Layers version D Example without fcn/convolutional layers at end
"""
with variable_scope.variable_scope(scope, 'vgg_16', [inputs]) as sc:
net = layers_lib.repeat(inputs, 2, layers.conv2d, 64, [3, 3], scope='conv1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net, 2, layers.conv2d, 128, [3, 3], scope='conv2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net, 3, layers.conv2d, 256, [3, 3], scope='conv3')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net, 3, layers.conv2d, 512, [3, 3], scope='conv4')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net, 3, layers.conv2d, 512, [3, 3], scope='conv5')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
net = layers.conv2d(net, 4096, [7, 7], padding='VALID', scope='fc6')
if middleRepr:
return net
net = tf.layers.dense(net, embedding_size,
activation=tf.nn.relu, kernel_initializer=xav_init(), name='fc8')
return net
def _embeddings(self, pretrained=False, scope_name=None):
"""Compute word embeddings for sentence.
Parameters
----------
pretrained: bool, default False
Whether to use pretrained embeddings
scope_name: str, default None
Variable scope
"""
if not scope_name:
scope_name = "Embedding"
self.sentence_ph = tf.placeholder(dtype=tf.int32, shape=[None, self.time_steps + 1],
name="Sentence_placeholder")
with tf.variable_scope(scope_name, reuse=tf.AUTO_REUSE):
self.embedding_matrix = tf.get_variable(
name="embedding_matrix",
shape=[self.len_corpus, self.embedding_size],
initializer=xav_init()
)
if pretrained:
print("Loading pretrained embeddings...")
load_embedding(session=self.session,
vocab=self.dataset.word_to_idx,
emb=self.embedding_matrix,
path=self.dataset.embedding_file,
vocab_size=self.len_corpus,
dim_embedding=self.embedding_size)
self.word_embeddings = tf.nn.embedding_lookup(self.embedding_matrix,
self.sentence_ph)
def _output_layer(self, scope_name=None):
"""Self explanatory."""
if scope_name is None:
scope_name = "Output_layer"
if self.project:
shape = [self.project_size, self.len_corpus]
else:
shape = [self.lstm_hidden_size, self.len_corpus]
with tf.variable_scope(scope_name, reuse=tf.AUTO_REUSE):
self.output_layer = dict()
self.output_layer['weights'] = tf.get_variable(name="weights",
shape=shape, dtype=tf.float32, initializer=xav_init())
self.output_layer['bias'] = tf.get_variable(name='bias',
shape=[self.len_corpus], dtype=tf.float32, initializer=xav_init()
)
def _projection_layer(self, scope_name=None):
"""Creates the weight matrix for projection, when a larger LSTM is used."""
if scope_name is None:
scope_name = "Projection"
with tf.variable_scope(scope_name, reuse=tf.AUTO_REUSE):
self.project_W = tf.get_variable(name="proj_weights",
shape=[self.lstm_hidden_size, self.project_size],
dtype=tf.float32,
initializer=xav_init()
)