def residual(self, inputs, filters, bottleneck=False, sub_sampling=False):
strides = 2 if sub_sampling else 1
kernel_size = (1, 1) if bottleneck else (3, 3)
outputs = batch_normalization()(inputs)
outputs = relu()(outputs)
outputs = conv2d(filters=filters,
kernel_size=kernel_size,
strides=strides)(outputs)
if sub_sampling or bottleneck:
# inputs = batch_normalization()(inputs)
# inputs = relu()(inputs)
inputs = conv2d(filters=4 * filters if bottleneck else filters,
kernel_size=(1, 1),
strides=strides)(inputs)
outputs = batch_normalization()(outputs)
outputs = relu()(outputs)
outputs = conv2d(filters=filters, kernel_size=(3, 3),
strides=1)(outputs)
if bottleneck:
outputs = batch_normalization()(outputs)
outputs = relu()(outputs)
outputs = conv2d(filters=4 * filters,
kernel_size=kernel_size,
strides=1)(outputs)
outputs = add()([inputs, outputs])
return outputs
def build(self, input_shape, classes=10):
filters = 16
inputs = keras.Input(shape=input_shape)
outputs = batch_normalization()(inputs)
outputs = relu()(outputs)
outputs = conv2d(filters=filters, kernel_size=(3, 3))(outputs)
filters = filters * self.widening
for i in range(3):
for j in range(self.blocks):
down_sampling = True if (i > 0 and j == 0) else False
outputs = self.residual(outputs,
filters=filters,
down_sampling=down_sampling)
filters = 2 * filters
outputs = batch_normalization()(outputs)
outputs = relu()(outputs)
outputs = average_pooling2d(pool_size=(8, 8))(outputs)
outputs = dense(10)(outputs)
outputs = softmax()(outputs)
model = keras.Model(inputs, outputs)
model.summary()
return model
def conv_block(self, inputs):
outputs = batch_normalization()(inputs)
if self.bottleneck:
outputs = relu()(outputs)
outputs = conv2d(filters=4 * self.growth_rate,
kernel_size=(1, 1))(outputs)
outputs = batch_normalization()(outputs)
outputs = relu()(outputs)
outputs = conv2d(filters=self.growth_rate, kernel_size=(3, 3))(outputs)
outputs = concat()([inputs, outputs])
return outputs
def Dense_net(self, input_x):
x = layers.conv2d(input_x, filters=2*self.filters, kernel_size=[7, 7], strides=[2, 2],
kernel_regularizer=layers.l2_regularizer(0.0005),
padding='valid', activation=None, name='conv0')
x = self.dense_block(input_x=x, nb_layers=6, layer_name='dense_1')
x = self.transition_layer(x, scope='trans_1')
x = self.dense_block(input_x=x, nb_layers=12, layer_name='dense_2')
x = self.transition_layer(x, scope='trans_2')
x = self.dense_block(input_x=x, nb_layers=48, layer_name='dense_3')
x = self.transition_layer(x, scope='trans_3')
x = self.dense_block(input_x=x, nb_layers=32, layer_name='dense_final')
# 100 Layer
x = layers.batch_normalization(x, training=self.training, name='linear_batch')
x = layers.selu(x)
x = layers.global_ave_pool2d(x)
# x = flatten(x)
x = layers.fully_connected(x, self.class_num, use_bias=False, activation_fn=None, trainable=self.training,
name='full_connecting')
# x = tf.reshape(x, [-1, 10])
return x
def conv(self, inputs, filters, kernel_size, strides=1):
inputs = batch_normalization()(inputs)
inputs = relu()(inputs)
inputs = conv2d(filters=filters,
kernel_size=kernel_size,
strides=strides)(inputs)
return inputs
def bottleneck_layer(self, x, scope):
with tf.variable_scope(scope):
x = layers.batch_normalization(x, training=self.training, name=scope + '_batch1')
x = layers.selu(x)
x = layers.conv2d(x, filters=4 * self.filters, kernel_size=[1, 1], strides=[1, 1],
kernel_regularizer=layers.l2_regularizer(0.0005),
padding='same', activation=None, name=scope + '_conv1')
x = layers.drop_out(x, rate=self.dropout, training=self.training)
x = layers.batch_normalization(x, training=self.training, name=scope + '_batch2')
x = layers.selu(x)
x = layers.conv2d(x, filters=self.filters, kernel_size=[3, 3], strides=[1, 1],
kernel_regularizer=layers.l2_regularizer(0.0005),
padding='same', activation=None, name=scope + '_conv2')
x = layers.drop_out(x, rate=self.dropout, training=self.training)
return x
def transition_block(self, inputs):
filters = keras.backend.int_shape(inputs)[
self.batch_normalization_axis]
if self.compression:
filters = int(filters * (1 - self.reduction_rate))
outputs = batch_normalization()(inputs)
outputs = relu()(outputs)
outputs = conv2d(filters=filters, kernel_size=(1, 1))(outputs)
outputs = average_pooling2d(pool_size=(2, 2), strides=2)(outputs)
return outputs
def __init__(self,
input_layers=None,
outputs_layers=None,
inception_layers=None):
self.inception_version = 1
if input_layers is None:
self.input_layers = [
conv2d(filters=64,
kernel_size=(7, 7),
strides=2,
padding='same'),
max_pooling2d(pool_size=(3, 3), strides=2, padding='same'),
batch_normalization(),
conv2d(filters=192,
kernel_size=(1, 1),
strides=1,
padding='valid'),
conv2d(filters=192,
kernel_size=(3, 3),
strides=1,
padding='same'),
batch_normalization(),
max_pooling2d(pool_size=(3, 3), strides=2, padding='same')
]
else:
self.input_layers = input_layers
if outputs_layers is None:
self.output_layers = [
average_pooling2d(pool_size=(7, 7), strides=1,
padding='valid'),
flatten(),
dropout(0.4),
dense(1000),
softmax()
]
else:
self.output_layers = outputs_layers
def __init__(
self,
blocks,
filters,
bottleneck=False,
input_layers=[
batch_normalization(),
relu(),
conv2d(filters=64, kernel_size=(7, 7), strides=2),
],
output_layers=[average_pooling2d(pool_size=(2, 2)),
flatten()]):
self.blocks = blocks
self.filters = filters
self.bottleneck = bottleneck
self.bn_axis = -1 if keras.backend.image_data_format(
) == 'channels_last' else 1
self.input_layers = input_layers
self.output_layer = output_layers
def build(self, input_shape, classes):
# input
inputs = keras.Input(shape=input_shape)
filters = 64
outputs = conv2d(filters=filters,
kernel_size=(7, 7),
strides=2,
padding='same')(inputs)
outputs = max_pooling2d(pool_size=(3, 3), strides=2,
padding='same')(outputs)
filters = 192
# using batch norm instead of local response norm
outputs = batch_normalization()(outputs)
outputs = conv2d(filters=filters,
kernel_size=(1, 1),
strides=1,
padding='valid')(outputs)
outputs = conv2d(filters=filters,
kernel_size=(3, 3),
strides=1,
padding='same')(outputs)
# using batch norm instead of local response norm
outputs = batch_normalization()(outputs)
outputs = max_pooling2d(pool_size=(3, 3), strides=2,
padding='same')(outputs)
# inception (3a)
filters = 256
outputs = self.inception(inputs=outputs, filters=filters)
# inception (3b)
filters = 480
outputs = self.inception(inputs=outputs, filters=filters)
outputs = max_pooling2d(pool_size=(3, 3), strides=2,
padding='same')(outputs)
# inception (4a)
filters = 512
outputs = self.inception(inputs=outputs, filters=filters)
# if K.learning_phase() == 1:
outputs2 = self.classifier_aux(outputs, classes=classes)
# inception (4b)
outputs = self.inception(inputs=outputs, filters=filters)
# inception (4c)
outputs = self.inception(inputs=outputs, filters=filters)
# inception (4d)
filters = 528
outputs = self.inception(inputs=outputs, filters=filters)
# if K.learning_phase() == 1:
outputs1 = self.classifier_aux(outputs, classes=classes)
# inception (4e)
filters = 832
outputs = self.inception(inputs=outputs, filters=filters)
outputs = max_pooling2d(pool_size=(2, 2), strides=2,
padding='same')(outputs)
# inception (5a)
outputs = self.inception(inputs=outputs, filters=filters)
# inception (5b)
filters = 1024
outputs = self.inception(inputs=outputs, filters=filters)
# classifier
outputs0 = self.classifier_main(inputs=outputs, classes=1000)
model = keras.Model(inputs=inputs,
outputs=[outputs0, outputs1, outputs2])
model.summary()
return model
def separable_conv_block(inputs, filters, kernel_size=(3, 3), strides=1):
inputs = separable_conv2d(filters=filters, kernel_size=kernel_size)(inputs)
inputs = batch_normalization()(inputs)
return inputs
def build_encoder(x):
# The encoder uses the deep residual network.
outputs = []
pooling = [1, 2, 2, 1]
shape = x.get_shape().as_list()
bs = shape[0]
seq = shape[1]
temp_shape = [bs * seq] + shape[2:]
x = tf.reshape(x, temp_shape)
# print x.get_shape().as_list()
# layer 0
with tf.variable_scope("encoder_layer0", reuse=tf.AUTO_REUSE):
conv0_0 = layers.conv_layer(name="conv0_0",
x=x,
filter_shape=layers.create_variable(
"filter0_0", shape=[7, 7, 3, 96]))
conv0_0 = layers.batch_normalization(conv0_0, "conv0_0_bn")
conv0_0 = layers.relu_layer(conv0_0)
conv0_1 = layers.conv_layer(name="conv0_1",
x=conv0_0,
filter_shape=layers.create_variable(
"filter0_1", shape=[3, 3, 96, 96]))
conv0_1 = layers.batch_normalization(conv0_1, "conv0_1_bn")
conv0_1 = layers.relu_layer(conv0_1)
shortcut0 = layers.conv_layer(name="shortcut",
x=x,
filter_shape=layers.create_variable(
"filter0_2", shape=[1, 1, 3, 96]))
shortcut0 = layers.batch_normalization(shortcut0, "shortcut0_bn")
shortcut0 = layers.relu_layer(shortcut0)
layer0 = layers.pooling_layer("pooling", conv0_1 + shortcut0, pooling)
outputs.append(layer0) # [bs * size, 64, 64, 96]
# layer 1
with tf.variable_scope("encoder_layer1", reuse=tf.AUTO_REUSE):
conv1_0 = layers.conv_layer(name="conv1_0",
x=layer0,
filter_shape=layers.create_variable(
"filter1_0", shape=[3, 3, 96, 128]))
conv1_0 = layers.batch_normalization(conv1_0, "conv1_0_bn")
conv1_0 = layers.relu_layer(conv1_0)
conv1_1 = layers.conv_layer(name="conv1_1",
x=conv1_0,
filter_shape=layers.create_variable(
"filter1_1", shape=[3, 3, 128, 128]))
conv1_1 = layers.batch_normalization(conv1_1, "conv1_1_bn")
conv1_1 = layers.relu_layer(conv1_1)
shortcut1 = layers.conv_layer(name="shortcut",
x=layer0,
filter_shape=layers.create_variable(
"filter1_2", shape=[1, 1, 96, 128]))
shortcut1 = layers.batch_normalization(shortcut1, "shortcut1_bn")
shortcut1 = layers.relu_layer(shortcut1)
layer1 = layers.pooling_layer("pooling", conv1_1 + shortcut1, pooling)
outputs.append(layer1) # [bs * size, 32, 32, 128]
# layer 2
with tf.variable_scope("encoder_layer2", reuse=tf.AUTO_REUSE):
conv2_0 = layers.conv_layer(name="conv2_0",
x=layer1,
filter_shape=layers.create_variable(
"filter2_0", shape=[3, 3, 128, 256]))
conv2_0 = layers.batch_normalization(conv2_0, "conv2_0_bn")
conv2_0 = layers.relu_layer(conv2_0)
conv2_1 = layers.conv_layer(name="conv2_1",
x=conv2_0,
filter_shape=layers.create_variable(
"filter2_1", shape=[3, 3, 256, 256]))
conv2_1 = layers.batch_normalization(conv2_1, "conv2_1_bn")
conv2_1 = layers.relu_layer(conv2_1)
shortcut2 = layers.conv_layer(name="shortcut",
x=layer1,
filter_shape=layers.create_variable(
"filter2_2", shape=[1, 1, 128, 256]))
shortcut2 = layers.batch_normalization(shortcut2, "shortcut2_bn")
shortcut2 = layers.relu_layer(shortcut2)
layer2 = layers.pooling_layer("pooling", conv2_1 + shortcut2, pooling)
outputs.append(layer2) # [bs * size, 16, 16, 256]
# layer 3
with tf.variable_scope("encoder_layer3", reuse=tf.AUTO_REUSE):
conv3_0 = layers.conv_layer(name="conv3_0",
x=layer2,
filter_shape=layers.create_variable(
"filter3_0", shape=[3, 3, 256, 256]))
conv3_0 = layers.batch_normalization(conv3_0, "conv3_0_bn")
conv3_0 = layers.relu_layer(conv3_0)
conv3_1 = layers.conv_layer(name="conv3_1",
x=conv3_0,
filter_shape=layers.create_variable(
"filter3_1", shape=[3, 3, 256, 256]))
conv3_1 = layers.batch_normalization(conv3_1, "conv3_1_bn")
conv3_1 = layers.relu_layer(conv3_1)
layer3 = layers.pooling_layer("pooling", conv3_1, pooling)
outputs.append(layer3) # [bs * size, 8, 8, 256]
# layer 4
with tf.variable_scope("encoder_layer4", reuse=tf.AUTO_REUSE):
conv4_0 = layers.conv_layer(name="conv4_0",
x=layer3,
filter_shape=layers.create_variable(
"filter4_0", shape=[3, 3, 256, 256]))
conv4_0 = layers.batch_normalization(conv4_0, "conv4_0_bn")
conv4_0 = layers.relu_layer(conv4_0)
conv4_1 = layers.conv_layer(name="conv4_1",
x=conv4_0,
filter_shape=layers.create_variable(
"filter4_1", shape=[3, 3, 256, 256]))
conv4_1 = layers.batch_normalization(conv4_1, "conv4_1_bn")
conv4_1 = layers.relu_layer(conv4_1)
shortcut4 = layers.conv_layer(name="shortcut",
x=layer3,
filter_shape=layers.create_variable(
"filter4_2", shape=[1, 1, 256, 256]))
shortcut4 = layers.batch_normalization(shortcut4, "shortcut4_bn")
shortcut4 = layers.relu_layer(shortcut4)
layer4 = layers.pooling_layer("pooling", conv4_1 + shortcut4, pooling)
outputs.append(layer4) # [bs * size, 4, 4, 256]
# layer 5
with tf.variable_scope("encoder_layer5", reuse=tf.AUTO_REUSE):
conv5_0 = layers.conv_layer(name="conv5_0",
x=layer4,
filter_shape=layers.create_variable(
"filter5_0", shape=[3, 3, 256, 256]))
conv5_0 = layers.batch_normalization(conv5_0, "conv5_0_bn")
conv5_0 = layers.relu_layer(conv5_0)
conv5_1 = layers.conv_layer(name="conv5_1",
x=conv5_0,
filter_shape=layers.create_variable(
"filter5_1", shape=[3, 3, 256, 256]))
conv5_1 = layers.batch_normalization(conv5_1, "conv5_1_bn")
conv5_1 = layers.relu_layer(conv5_1)
shortcut5 = layers.conv_layer(name="shortcut",
x=layer4,
filter_shape=layers.create_variable(
"filter5_2", shape=[1, 1, 256, 256]))
shortcut5 = layers.batch_normalization(shortcut5, "shortcut5_bn")
shortcut5 = layers.relu_layer(shortcut5)
layer5 = layers.pooling_layer("pooling", conv5_1 + shortcut5, pooling)
outputs.append(layer5) # [bs * size, 2, 2, 256]
final_shape = [bs, seq, fc_layer_size[0]]
# Flatten layer and fully connected layer
flatten = layers.flatten_layer(layer5)
outputs.append(flatten)
with tf.variable_scope("fc_layer", reuse=tf.AUTO_REUSE):
layer_fc = layers.fully_connected_layer(flatten, fc_layer_size[0],
"fclayer_w", "fclayer_b")
# layer_fc = layers.batch_normalization(layer_fc, "fc_bn")
layer_fc = layers.relu_layer(layer_fc)
outputs.append(layer_fc) # [bs * size, 1024]
# [bs, size, 1024]
return tf.reshape(outputs[-1], final_shape)