def _declare_variables(self):
with tf.variable_scope('vfeedbacknet_{}'.format(Model.model_name)):
with tf.variable_scope('feedbackcell1'):
self.feedbackLSTMCell1 = FeedbackLSTMCell_stack1(
[14, 14, 128], Model.NFEEDBACK)
with tf.variable_scope('convlstm1'):
with tf.variable_scope('rnn'):
with tf.variable_scope('conv_lstm_cell'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
n = 128
m = 4 * n
input_size = [7, 7, n]
kernel2d_size = [3, 3]
kernel_size = kernel2d_size + [2 * n] + [m]
with tf.variable_scope('convlstm'):
kernel = tf.get_variable('kernel',
kernel_size,
initializer=initializer,
regularizer=regularizer)
W_ci = tf.get_variable('W_ci',
input_size,
initializer=initializer,
regularizer=regularizer)
W_cf = tf.get_variable('W_cf',
input_size,
initializer=initializer,
regularizer=regularizer)
W_co = tf.get_variable('W_co',
input_size,
initializer=initializer,
regularizer=regularizer)
bias = tf.get_variable(
'bias', [m],
initializer=tf.zeros_initializer(),
regularizer=regularizer)
self.convLSTMCell1 = ConvLSTMCell(input_size[:2], n, [3, 3])
with tf.variable_scope('reshape_convs'):
with tf.variable_scope('conv1'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
kernel = tf.get_variable('kernel',
shape=[7, 7, 3, 32],
dtype=tf.float32,
regularizer=regularizer,
initializer=initializer)
biases = tf.get_variable('biases',
shape=[32],
dtype=tf.float32,
regularizer=regularizer,
initializer=initializer)
with tf.variable_scope('conv2'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
kernel = tf.get_variable('kernel',
shape=[3, 3, 32, 64],
dtype=tf.float32,
regularizer=regularizer,
initializer=initializer)
biases = tf.get_variable('biases',
shape=[64],
dtype=tf.float32,
regularizer=regularizer,
initializer=initializer)
with tf.variable_scope('conv3'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
kernel = tf.get_variable('kernel',
shape=[3, 3, 64, 128],
dtype=tf.float32,
regularizer=regularizer,
initializer=initializer)
biases = tf.get_variable('biases',
shape=[128],
dtype=tf.float32,
regularizer=regularizer,
initializer=initializer)
# with tf.variable_scope('conv4'):
# regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
# initializer = tf.contrib.layers.xavier_initializer()
# kernel = tf.get_variable('kernel', shape=[3, 3, 512, 1024], dtype=tf.float32, regularizer=regularizer, initializer=initializer)
# biases = tf.get_variable('biases', shape=[1024], dtype=tf.float32, regularizer=regularizer, initializer=initializer)
# with tf.variable_scope('conv5'):
# regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
# initializer = tf.contrib.layers.xavier_initializer()
# kernel = tf.get_variable('kernel', shape=[3, 3, 128, 256], dtype=tf.float32, regularizer=regularizer, initializer=initializer)
# biases = tf.get_variable('biases', shape=[256], dtype=tf.float32, regularizer=regularizer, initializer=initializer)
with tf.variable_scope('fc1'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
trainable = False if self.train_fc == 'NO' else True
weight = tf.get_variable('weights',
shape=[128, 128],
dtype=tf.float32,
initializer=initializer,
regularizer=regularizer,
trainable=trainable)
biases = tf.get_variable('biases',
shape=[128],
dtype=tf.float32,
initializer=initializer,
regularizer=regularizer,
trainable=trainable)
with tf.variable_scope('fc2'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
trainable = False if self.train_fc == 'NO' else True
weight = tf.get_variable('weights',
shape=[128, self.num_classes],
dtype=tf.float32,
initializer=initializer,
regularizer=regularizer,
trainable=trainable)
biases = tf.get_variable('biases',
shape=[self.num_classes],
dtype=tf.float32,
initializer=initializer,
regularizer=regularizer,
trainable=trainable)
def _declare_variables(self):
with tf.variable_scope('vfeedbacknet_model1'):
with tf.variable_scope('feedbackcell1'):
self.feedbackLSTMCell1 = FeedbackLSTMCell_stack1(
[4, 4, 512], Model.NFEEDBACK)
with tf.variable_scope('convlstm1'):
with tf.variable_scope('rnn'):
with tf.variable_scope('conv_lstm_cell'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
n = 512
m = 4 * n
input_size = [4, 4, n]
kernel2d_size = [3, 3]
kernel_size = kernel2d_size + [2 * n] + [m]
with tf.variable_scope('convlstm'):
kernel = tf.get_variable('kernel',
kernel_size,
initializer=initializer,
regularizer=regularizer)
W_ci = tf.get_variable('W_ci',
input_size,
initializer=initializer,
regularizer=regularizer)
W_cf = tf.get_variable('W_cf',
input_size,
initializer=initializer,
regularizer=regularizer)
W_co = tf.get_variable('W_co',
input_size,
initializer=initializer,
regularizer=regularizer)
bias = tf.get_variable(
'bias', [m],
initializer=tf.zeros_initializer(),
regularizer=regularizer)
self.convLSTMCell1 = ConvLSTMCell([4, 4], 512, [3, 3])
with tf.variable_scope('convlstm2'):
with tf.variable_scope('rnn'):
with tf.variable_scope('conv_lstm_cell'):
regularizer = None # tf.contrib.layers.l2_regularizer(scale=0.25)
initializer = tf.contrib.layers.xavier_initializer()
n = 512
m = 4 * n
input_size = [4, 4, n]
kernel2d_size = [3, 3]
kernel_size = kernel2d_size + [2 * n] + [m]
with tf.variable_scope('convlstm'):
kernel = tf.get_variable('kernel',
kernel_size,
initializer=initializer,
regularizer=regularizer)
W_ci = tf.get_variable('W_ci',
input_size,
initializer=initializer,
regularizer=regularizer)
W_cf = tf.get_variable('W_cf',
input_size,
initializer=initializer,
regularizer=regularizer)
W_co = tf.get_variable('W_co',
input_size,
initializer=initializer,
regularizer=regularizer)
bias = tf.get_variable(
'bias', [m],
initializer=tf.zeros_initializer(),
regularizer=regularizer)
self.convLSTMCell2 = ConvLSTMCell([4, 4], 512, [3, 3])
