def add_residual_dense_block(in_layer, filter_dims, num_layers, act_func=tf.nn.relu, bn_phaze=False, scope='residual_dense_block'):
with tf.variable_scope(scope):
l = in_layer
input_dims = in_layer.get_shape().as_list()
num_channel_in = input_dims[-1]
num_channel_out = input_dims[-1]
for i in range(num_layers):
l = layers.add_dense_layer(l, filter_dims=filter_dims, act_func=act_func, bn_phaze=bn_phaze,
scope='layer' + str(i))
l = layers.add_dense_transition_layer(l, filter_dims=[1, 1, num_channel_out], act_func=act_func,
scope='dense_transition_1', bn_phaze=bn_phaze, use_pool=False)
l = tf.add(l, in_layer)
return l
def cnn_network(x, activation='swish', scope='cnn_network', reuse=False, bn_phaze=False, keep_prob=0.5):
with tf.variable_scope(scope):
if reuse:
tf.get_variable_scope().reuse_variables()
if activation == 'swish':
act_func = util.swish
elif activation == 'relu':
act_func = tf.nn.relu
else:
act_func = tf.nn.sigmoid
l = layers.conv(x, scope='conv1', filter_dims=[5, 5, 256], stride_dims=[1, 1], non_linear_fn=None, bias=False)
with tf.variable_scope('dense_block_1'):
for i in range(g_dense_block_layers):
l = layers.add_dense_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func, bn_phaze=bn_phaze,
scope='layer' + str(i))
l = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth], act_func=act_func,
scope='dense_transition_1', bn_phaze=bn_phaze)
l = add_residual_dense_block(l, filter_dims=[3, 3, g_dense_block_depth], num_layers=3,
act_func=act_func, bn_phaze=bn_phaze, scope='block_1')
l = tf.nn.max_pool(l, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
l = add_residual_dense_block(l, filter_dims=[3, 3, g_dense_block_depth], num_layers=3,
act_func=act_func, bn_phaze=bn_phaze, scope='block_2')
l = tf.nn.max_pool(l, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
l = add_residual_dense_block(l, filter_dims=[3, 3, g_dense_block_depth], num_layers=3,
act_func=act_func, bn_phaze=bn_phaze, scope='block_3')
l = tf.nn.max_pool(l, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
l = add_residual_dense_block(l, filter_dims=[3, 3, g_dense_block_depth], num_layers=3,
act_func=act_func, bn_phaze=bn_phaze, scope='block_4')
l = add_residual_dense_block(l, filter_dims=[3, 3, g_dense_block_depth], num_layers=3,
act_func=act_func, bn_phaze=bn_phaze, scope='block_5')
with tf.variable_scope('dense_block_last'):
for i in range(g_dense_block_layers):
l = layers.add_dense_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer' + str(i))
scale_layer = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth],
act_func=act_func,
scope='dense_transition_1', bn_phaze=bn_phaze,
use_pool=False)
last_dense_layer = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth],
act_func=act_func,
scope='dense_transition_2', bn_phaze=bn_phaze,
use_pool=False)
scale_layer = act_func(scale_layer)
last_dense_layer = act_func(last_dense_layer)
'''
with tf.variable_scope('residual_block_1'):
r = layers.add_residual_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer1')
r = layers.add_residual_layer(r, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer2')
l = tf.add(l, r)
with tf.variable_scope('dense_block_2'):
for i in range(g_dense_block_layers):
l = layers.add_dense_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func, bn_phaze=bn_phaze,
scope='layer' + str(i))
l = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth], act_func=act_func,
scope='dense_transition_1', bn_phaze=bn_phaze)
with tf.variable_scope('residual_block_2'):
r = layers.add_residual_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer1')
r = layers.add_residual_layer(r, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer2')
l = tf.add(l, r)
with tf.variable_scope('dense_block_3'):
for i in range(g_dense_block_layers):
l = layers.add_dense_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func, bn_phaze=bn_phaze,
scope='layer' + str(i))
l = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth], act_func=act_func,
scope='dense_transition_1', bn_phaze=bn_phaze)
with tf.variable_scope('residual_block_3'):
r = layers.add_residual_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer1')
r = layers.add_residual_layer(r, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer2')
l = tf.add(l, r)
with tf.variable_scope('dense_block_4'):
for i in range(g_dense_block_layers):
l = layers.add_dense_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func, bn_phaze=bn_phaze,
scope='layer' + str(i))
l = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth], act_func=act_func,
scope='dense_transition_1', bn_phaze=bn_phaze)
with tf.variable_scope('residual_block_4'):
r = layers.add_residual_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer1')
r = layers.add_residual_layer(r, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer2')
l = tf.add(l, r)
with tf.variable_scope('dense_block_last'):
for i in range(g_dense_block_layers):
l = layers.add_dense_layer(l, filter_dims=[3, 3, g_dense_block_depth], act_func=act_func,
bn_phaze=bn_phaze, scope='layer' + str(i))
scale_layer = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth], act_func=act_func,
scope='dense_transition_1', bn_phaze=bn_phaze, use_pool=False)
last_dense_layer = layers.add_dense_transition_layer(l, filter_dims=[1, 1, g_dense_block_depth],
act_func=act_func,
scope='dense_transition_2', bn_phaze=bn_phaze, use_pool=False)
scale_layer = act_func(scale_layer)
last_dense_layer = act_func(last_dense_layer)
'''
return last_dense_layer, scale_layer