深度学习--Inception-ResNet-v1网络结构

　　Inception V4的网络结构如下：　　

　　从图中可以看出，输入部分与V1到V3的输入部分有较大的差别，这样设计的目的为了：使用并行结构、不对称卷积核结构，可以在保证信息损失足够小的情况下，降低计算量。结构中1*1的卷积核也用来降维，并且也增加了非线性。
　　Inception-ResNet-v2与Inception-ResNet-v1的结构类似，除了stem部分。Inception-ResNet-v2的stem与V4的结构类似，Inception-ResNet-v2的输出chnnel要高。Reduction-A相同，Inception-ResNet-A、Inception-ResNet-B、Inception-ResNet-C和Reduction-B的结构与v1的类似，只不过输出的channel数量更多。
　　
　　Inception-ResNet-v1的总体网络结构如下所示：

Inception-ResNet-v1的Stem与V3的结构是一致的。
　　接下来主要说一下Inception-ResNet-v1的网络结构及代码的实现部分。

Stem结构

　　stem结构与V3的Stem结构类似。
对应的代码为

with slim.arg_scope([slim.conv2d, slim.max_pool2d, slim.avg_pool2d],stride=1, padding='SAME'):# 149 x 149 x 32net = slim.conv2d(inputs, 32, 3, stride=2, padding='VALID', scope='Conv2d_1a_3x3')end_points['Conv2d_1a_3x3'] = net# 147 x 147 x 32net = slim.conv2d(net, 32, 3, padding='VALID',
scope='Conv2d_2a_3x3')end_points['Conv2d_2a_3x3'] = net# 147 x 147 x 64net = slim.conv2d(net, 64, 3, scope='Conv2d_2b_3x3')
end_points['Conv2d_2b_3x3'] = net# 73 x 73 x 64net = slim.max_pool2d(net, 3, stride=2, padding='VALID', scope='MaxPool_3a_3x3')end_points['MaxPool_3a_3x3'] = net# 73 x 73 x 80net = slim.conv2d(net, 80, 1, padding='VALID',
scope='Conv2d_3b_1x1')end_points['Conv2d_3b_1x1'] = net# 71 x 71 x 192net = slim.conv2d(net, 192, 3, padding='VALID',
scope='Conv2d_4a_3x3')end_points['Conv2d_4a_3x3'] = net# 35 x 35 x 256net = slim.conv2d(net, 256, 3, stride=2, padding='VALID',
scope='Conv2d_4b_3x3')end_points['Conv2d_4b_3x3'] = net

Inception-resnet-A模块

　　Inception-resnet-A模块是要重复5次的，网络结构为：

　　对应的代码表示为：

# Inception-Renset-A
def block35(net, scale=1.0, activation_fn=tf.nn.relu, scope=None, reuse=None):"""Builds the 35x35 resnet block."""with tf.variable_scope(scope, 'Block35', [net], reuse=reuse):with tf.variable_scope('Branch_0'):# 35 × 35 × 32tower_conv = slim.conv2d(net, 32, 1, scope='Conv2d_1x1')with tf.variable_scope('Branch_1'):# 35 × 35 × 32tower_conv1_0 = slim.conv2d(net, 32, 1, scope='Conv2d_0a_1x1')# 35 × 35 × 32tower_conv1_1 = slim.conv2d(tower_conv1_0, 32, 3, scope='Conv2d_0b_3x3')with tf.variable_scope('Branch_2'):# 35 × 35 × 32tower_conv2_0 = slim.conv2d(net, 32, 1, scope='Conv2d_0a_1x1')# 35 × 35 × 32tower_conv2_1 = slim.conv2d(tower_conv2_0, 32, 3, scope='Conv2d_0b_3x3')# 35 × 35 × 32tower_conv2_2 = slim.conv2d(tower_conv2_1, 32, 3, scope='Conv2d_0c_3x3')# 35 × 35 × 96mixed = tf.concat([tower_conv, tower_conv1_1, tower_conv2_2], 3)# 35 × 35 × 256up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,activation_fn=None, scope='Conv2d_1x1')# 使用残差网络scale = 0.17net += scale * upif activation_fn:net = activation_fn(net)return net# 5 x Inception-resnet-A
net = slim.repeat(net, 5, block35, scale=0.17)
end_points['Mixed_5a'] = net

Reduction-A结构

　　Reduction-A中含有4个参数k、l、 m、 n，它们对应的值分别为：192, 192, 256, 384，在该层网络结构，输入为35×35×256，输出为17×17×896.

def reduction_a(net, k, l, m, n):# 192, 192, 256, 384with tf.variable_scope('Branch_0'):# 17×17×384tower_conv = slim.conv2d(net, n, 3, stride=2, padding='VALID',scope='Conv2d_1a_3x3')with tf.variable_scope('Branch_1'):# 35×35×192tower_conv1_0 = slim.conv2d(net, k, 1, scope='Conv2d_0a_1x1')# 35×35×192tower_conv1_1 = slim.conv2d(tower_conv1_0, l, 3,scope='Conv2d_0b_3x3')# 17×17×256tower_conv1_2 = slim.conv2d(tower_conv1_1, m, 3,stride=2, padding='VALID',scope='Conv2d_1a_3x3')with tf.variable_scope('Branch_2'):# 17×17×256tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',scope='MaxPool_1a_3x3')# 17×17×896net = tf.concat([tower_conv, tower_conv1_2, tower_pool], 3)return net# Reduction-A
with tf.variable_scope('Mixed_6a'):net = reduction_a(net, 192, 192, 256, 384)end_points['Mixed_6a'] = net

Inception-Resnet-B

　　Inception-Resnet-B模块是要重复10次，输入为17×17×896，输出为17×17×896，网络结构为：

# Inception-Renset-B
def block17(net, scale=1.0, activation_fn=tf.nn.relu, scope=None, reuse=None):"""Builds the 17x17 resnet block."""with tf.variable_scope(scope, 'Block17', [net], reuse=reuse):with tf.variable_scope('Branch_0'):# 17*17*128tower_conv = slim.conv2d(net, 128, 1, scope='Conv2d_1x1')with tf.variable_scope('Branch_1'):# 17*17*128tower_conv1_0 = slim.conv2d(net, 128, 1, scope='Conv2d_0a_1x1')# 17*17*128tower_conv1_1 = slim.conv2d(tower_conv1_0, 128, [1, 7],scope='Conv2d_0b_1x7')# 17*17*128tower_conv1_2 = slim.conv2d(tower_conv1_1, 128, [7, 1],scope='Conv2d_0c_7x1')# 17*17*256mixed = tf.concat([tower_conv, tower_conv1_2], 3)# 17*17*896up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,activation_fn=None, scope='Conv2d_1x1')net += scale * upif activation_fn:net = activation_fn(net)return net# 10 x Inception-Resnet-B
net = slim.repeat(net, 10, block17, scale=0.10)
end_points['Mixed_6b'] = net

Reduction-B

　　Reduction-B的输入为17*17*896，输出为8*8*1792。网络结构为：

对应的代码为：

def reduction_b(net):with tf.variable_scope('Branch_0'):# 17*17*256tower_conv = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')# 8*8*384tower_conv_1 = slim.conv2d(tower_conv, 384, 3, stride=2,padding='VALID', scope='Conv2d_1a_3x3')with tf.variable_scope('Branch_1'):# 17*17*256tower_conv1 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')# 8*8*256tower_conv1_1 = slim.conv2d(tower_conv1, 256, 3, stride=2,padding='VALID', scope='Conv2d_1a_3x3')with tf.variable_scope('Branch_2'):# 17*17*256tower_conv2 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')# 17*17*256tower_conv2_1 = slim.conv2d(tower_conv2, 256, 3,scope='Conv2d_0b_3x3')# 8*8*256tower_conv2_2 = slim.conv2d(tower_conv2_1, 256, 3, stride=2,padding='VALID', scope='Conv2d_1a_3x3')with tf.variable_scope('Branch_3'):# 8*8*896tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',scope='MaxPool_1a_3x3')# 8*8*1792net = tf.concat([tower_conv_1, tower_conv1_1,tower_conv2_2, tower_pool], 3)return net
# Reduction-B
with tf.variable_scope('Mixed_7a'):net = reduction_b(net)
end_points['Mixed_7a'] = net

Inception-Resnet-C结构

　　Inception-Resnet-C结构重复5次。它输入为8*8*1792，输出为8*8*1792。对应的结构为：

　　对应的代码为：

# Inception-Resnet-C
def block8(net, scale=1.0, activation_fn=tf.nn.relu, scope=None, reuse=None):"""Builds the 8x8 resnet block."""with tf.variable_scope(scope, 'Block8', [net], reuse=reuse):with tf.variable_scope('Branch_0'):# 8*8*192tower_conv = slim.conv2d(net, 192, 1, scope='Conv2d_1x1')with tf.variable_scope('Branch_1'):# 8*8*192tower_conv1_0 = slim.conv2d(net, 192, 1, scope='Conv2d_0a_1x1')# 8*8*192tower_conv1_1 = slim.conv2d(tower_conv1_0, 192, [1, 3],scope='Conv2d_0b_1x3')# 8*8*192tower_conv1_2 = slim.conv2d(tower_conv1_1, 192, [3, 1],scope='Conv2d_0c_3x1')# 8*8*384mixed = tf.concat([tower_conv, tower_conv1_2], 3)# 8*8*1792up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,activation_fn=None, scope='Conv2d_1x1')# scale=0.20net += scale * upif activation_fn:net = activation_fn(net)return net
# 5 x Inception-Resnet-C
net = slim.repeat(net, 5, block8, scale=0.20)
end_points['Mixed_8a'] = net

　　但是在facenet中，接下来又是一层Inception-Resnet-C，但是它没有重复，并且没有激活函数。输入与输出大小相同。

net = block8(net, activation_fn=None)
end_points['Mixed_8b'] = net

结果输出

　　结果输出包含Average Pooling和Dropout (keep 0.8)及Softmax三层，这里我们以facenet中为例：具体的代码如下：

with tf.variable_scope('Logits'):end_points['PrePool'] = net#pylint: disable=no-member# Average Pooling层，输出为8×8×1792net = slim.avg_pool2d(net, net.get_shape()[1:3], padding='VALID',scope='AvgPool_1a_8x8')#扁平除了batch_size维度的其它维度。使输出变为：[batch_size, ...]net = slim.flatten(net)#dropout层net = slim.dropout(net, dropout_keep_prob, is_training=is_training,scope='Dropout')end_points['PreLogitsFlatten'] = net# 全链接层。输出为batch_size×128net = slim.fully_connected(net, bottleneck_layer_size, activation_fn=None,scope='Bottleneck', reuse=False)

　　至此，inception_resnet_v1网络结构就结束了，但facenet的代码分析未完，待续～～～～