def bottleneck_block(self, input_x, group, name, is_first_block=False):
with tf.variable_scope(name):
# 1x1 convolution responsible for reducing the depth
with tf.variable_scope('conv_in'):
stride = 2 if is_first_block else 1
conv = conv_layer_3d(input_x, num_filters=group.bottleneck_size, add_reg=self.conf.L2_reg,
kernel_size=1, stride=stride, layer_name='CONV')
conv = relu(batch_normalization(conv, self.is_training, scope='BN'))
with tf.variable_scope('conv_bottleneck'):
conv = conv_layer_3d(conv, num_filters=group.bottleneck_size, kernel_size=3,
add_reg=self.conf.L2_reg, layer_name='CONV')
conv = relu(batch_normalization(conv, self.is_training, scope='BN'))
# 1x1 convolution responsible for increasing the depth
with tf.variable_scope('conv_out'):
conv = conv_layer_3d(conv, num_filters=group.out_filters, kernel_size=1,
add_reg=self.conf.L2_reg, layer_name='CONV')
conv = batch_normalization(conv, self.is_training, scope='BN')
# shortcut connections that turn the network into its counterpart
# residual function (identity shortcut)
with tf.variable_scope('shortcut'):
if is_first_block:
shortcut = conv_layer_3d(input_x, num_filters=group.out_filters, stride=2, kernel_size=1,
add_reg=self.conf.L2_reg, layer_name='CONV_shortcut')
shortcut = batch_normalization(shortcut, self.is_training, scope='BN_shortcut')
assert (shortcut.get_shape().as_list() == conv.get_shape().as_list()), \
"Tensor sizes of the two branches are not matched!"
res = shortcut + conv
else:
res = conv + input_x
assert (input_x.get_shape().as_list() == conv.get_shape().as_list()), \
"Tensor sizes of the two branches are not matched!"
return relu(res)
def build_network(self, x):
# Building network...
with tf.variable_scope('ResNet'):
net = conv_layer_3d(x, num_filters=64, kernel_size=4, stride=1, add_reg=self.conf.L2_reg,
layer_name='CONV0')
net = relu(batch_normalization(net, training=self.is_training, scope='BN1'))
# net = max_pool_3d(net, pool_size=3, stride=2, name='MaxPool0')
# Create the bottleneck groups, each of which contains `num_blocks` bottleneck blocks.
for group_i, group in enumerate(self.groups):
first_block = True
for block_i in range(group.num_blocks):
block_name = 'group_%d/block_%d' % (group_i, block_i)
net = self.bottleneck_block(net, group, block_name, is_first_block=first_block)
first_block = False
net = average_pool_3d(net, pool_size=2, stride=1, name='avg_pool')
net = flatten(net)
net = fc_layer(net, num_units=75, add_reg=self.conf.L2_reg, layer_name='Fc1')
net = dropout(net, self.conf.dropout_rate, training=self.is_training)
self.logits = fc_layer(net, num_units=self.conf.num_cls, add_reg=self.conf.L2_reg, layer_name='Fc2')
# [?, num_cls]
self.probs = tf.nn.softmax(self.logits)
# [?, num_cls]
self.y_pred = tf.to_int32(tf.argmax(self.probs, 1))
def build_network(self, x):
# Building network...
with tf.variable_scope('DenseNet'):
net = conv_layer(x,
num_filters=2 * self.k,
kernel_size=7,
stride=2,
layer_name='Conv0')
# net = max_pool(net, pool_size=3, stride=2, name='MaxPool0')
for l in range(self.conf.num_levels):
net = self.dense_block(net,
num_BBs=self.conf.num_BBs[l],
block_name='DB_' + str(l))
print('DB_{} shape: {}'.format(str(l + 1), net.get_shape()))
net = self.transition_layer(net, scope='TB_' + str(l))
print('TD_{} shape: {}'.format(str(l + 1), net.get_shape()))
# net = self.dense_block(net, num_BBs=32, block_name='Dense_final')
# print('DB_{} shape: {}'.format(str(l + 2), net.get_shape()))
net = batch_normalization(net,
training=self.is_training,
scope='BN_out')
net = relu(net)
net = global_average_pool(net)
net = flatten(net)
self.features = net
self.logits = fc_layer(net,
num_units=self.conf.num_cls,
add_reg=self.conf.L2_reg,
layer_name='Fc1')
# [?, num_cls]
self.prob = tf.nn.softmax(self.logits)
# [?, num_cls]
self.y_pred = tf.to_int32(tf.argmax(self.prob, 1))
def build_network(self, x):
# Building network...
with tf.variable_scope('CapsNet'):
net = batch_normalization(relu(conv_layer_3d(x, kernel_size=7, stride=2, num_filters=96,
add_reg=self.conf.L2_reg, layer_name='CONV1')),
training=self.is_training, scope='BN1')
net = max_pool_3d(net, pool_size=3, stride=2, padding='SAME', name='MaxPool1')
net = batch_normalization(relu(conv_layer_3d(net, kernel_size=5, stride=2, num_filters=256,
add_reg=self.conf.L2_reg, layer_name='CONV2')),
training=self.is_training, scope='BN2')
net = max_pool_3d(net, pool_size=3, stride=2, padding='SAME', name='MaxPool2')
net = batch_normalization(relu(conv_layer_3d(net, kernel_size=3, stride=1, num_filters=384,
add_reg=self.conf.L2_reg, layer_name='CONV3')),
training=self.is_training, scope='BN3')
net = batch_normalization(relu(conv_layer_3d(net, kernel_size=3, stride=1, num_filters=384,
add_reg=self.conf.L2_reg, layer_name='CONV4')),
training=self.is_training, scope='BN4')
net = batch_normalization(relu(conv_layer_3d(net, kernel_size=3, stride=1, num_filters=256,
add_reg=self.conf.L2_reg, layer_name='CONV5')),
training=self.is_training, scope='BN5')
net = max_pool_3d(net, pool_size=3, stride=2, padding='SAME', name='MaxPool3')
layer_flat = flatten(net)
net = relu(fc_layer(layer_flat, num_units=200, add_reg=self.conf.L2_reg, layer_name='FC1'))
net = dropout(net, self.conf.dropout_rate, training=self.is_training)
net = relu(fc_layer(net, num_units=75, add_reg=self.conf.L2_reg, layer_name='FC2'))
net = dropout(net, self.conf.dropout_rate, training=self.is_training)
self.features = net
self.logits = fc_layer(net, num_units=self.conf.num_cls, add_reg=self.conf.L2_reg, layer_name='FC3')
# [?, num_cls]
self.probs = tf.nn.softmax(self.logits)
# [?, num_cls]
self.y_pred = tf.to_int32(tf.argmax(self.probs, 1))
def bottleneck_block(self, x, scope):
with tf.variable_scope(scope):
x = batch_normalization(x, training=self.is_training, scope='BN1')
x = relu(x)
x = conv_layer(x,
num_filters=4 * self.k,
kernel_size=1,
layer_name='CONV1')
x = dropout(x,
rate=self.conf.dropout_rate,
training=self.is_training)
x = batch_normalization(x, training=self.is_training, scope='BN2')
x = relu(x)
x = conv_layer(x,
num_filters=self.k,
kernel_size=3,
layer_name='CONV2')
x = dropout(x,
rate=self.conf.dropout_rate,
training=self.is_training)
return x
def transition_layer(self, x, scope):
with tf.variable_scope(scope):
x = batch_normalization(x, training=self.is_training, scope='BN')
x = relu(x)
x = conv_layer(x,
num_filters=int(x.get_shape().as_list()[-1] *
self.conf.theta),
kernel_size=1,
layer_name='CONV')
x = dropout(x,
rate=self.conf.dropout_rate,
training=self.is_training)
x = average_pool(x, pool_size=2, stride=2, name='AVG_POOL')
return x
def build_network(self, x):
# Building network...
with tf.variable_scope('CapsNet'):
net = lrn(
relu(
conv_layer(x,
kernel_size=7,
stride=2,
num_filters=96,
trainable=self.conf.trainable,
add_reg=self.conf.L2_reg,
layer_name='CONV1')))
net = max_pool(net,
pool_size=3,
stride=2,
padding='SAME',
name='MaxPool1')
net = lrn(
relu(
conv_layer(net,
kernel_size=5,
stride=2,
num_filters=256,
trainable=self.conf.trainable,
add_reg=self.conf.L2_reg,
layer_name='CONV2')))
net = max_pool(net,
pool_size=3,
stride=2,
padding='SAME',
name='MaxPool2')
net = relu(
conv_layer(net,
kernel_size=3,
stride=1,
num_filters=384,
trainable=self.conf.trainable,
add_reg=self.conf.L2_reg,
layer_name='CONV3'))
net = relu(
conv_layer(net,
kernel_size=3,
stride=1,
num_filters=384,
trainable=self.conf.trainable,
add_reg=self.conf.L2_reg,
layer_name='CONV4'))
self.net_grad = relu(
conv_layer(net,
kernel_size=3,
stride=1,
num_filters=256,
trainable=self.conf.trainable,
add_reg=self.conf.L2_reg,
layer_name='CONV5'))
net = max_pool(self.net_grad,
pool_size=3,
stride=2,
padding='SAME',
name='MaxPool3')
layer_flat = flatten(net)
net = relu(
fc_layer(layer_flat,
num_units=512,
add_reg=self.conf.L2_reg,
trainable=self.conf.trainable,
layer_name='FC1'))
net = dropout(net,
self.conf.dropout_rate,
training=self.is_training)
net = relu(
fc_layer(net,
num_units=512,
add_reg=self.conf.L2_reg,
trainable=self.conf.trainable,
layer_name='FC2'))
net = dropout(net,
self.conf.dropout_rate,
training=self.is_training)
self.features = net
self.logits = fc_layer(net,
num_units=self.conf.num_cls,
add_reg=self.conf.L2_reg,
trainable=self.conf.trainable,
layer_name='FC3')
# [?, num_cls]
self.prob = tf.nn.softmax(self.logits)
# [?, num_cls]
self.y_pred = tf.to_int32(tf.argmax(self.prob, 1))