Inception V4的網路結構如下： 　　

　　從圖中可以看出，輸入部分與V1到V3的輸入部分有較大的差別，這樣設計的目的為了：使用並行結構、不對稱卷積核結構，可以在保證資訊損失足夠小的情況下，降低計算量。結構中1*1的卷積核也用來降維，並且也增加了非線性。
　　Inception-ResNet-v2與Inception-ResNet-v1的結構類似，除了stem部分。Inception-ResNet-v2的stem與V4的結構類似，Inception-ResNet-v2的輸出chnnel要高。Reduction-A相同，Inception-ResNet-A、Inception-ResNet-B、Inception-ResNet-C和Reduction-B的結構與v1的類似，只不過輸出的channel數量更多。
　　
　　Inception-ResNet-v1的總體網路結構如下所示：

Inception-ResNet-v1的Stem與V3的結構是一致的。
　　接下來主要說一下Inception-ResNet-v1的網路結構及程式碼的實現部分。

Stem結構

　　stem結構與V3的Stem結構類似。
對應的程式碼為

with slim.arg_scope([slim.conv2d, slim.max_pool2d, slim.avg_pool2d],stride=1, padding='SAME'):

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

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21
• 22
• 23
• 24
• 25
• 26
• 27

Inception-resnet-A模組

　　Inception-resnet-A模組是要重複5次的，網路結構為：

　　對應的程式碼表示為：

# 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 × 32
            tower_conv = slim.conv2d(net, 32, 1, scope='Conv2d_1x1')
        with tf.variable_scope('Branch_1'):
            # 35 × 35 × 32
            tower_conv1_0 = slim.conv2d(net, 32, 1, scope='Conv2d_0a_1x1')
            # 35 × 35 × 32
            tower_conv1_1 = slim.conv2d(tower_conv1_0, 32, 3, scope='Conv2d_0b_3x3')
        with tf.variable_scope('Branch_2'):
            # 35 × 35 × 32
            tower_conv2_0 = slim.conv2d(net, 32, 1, scope='Conv2d_0a_1x1')
            # 35 × 35 × 32
            tower_conv2_1 = slim.conv2d(tower_conv2_0, 32, 3, scope='Conv2d_0b_3x3')
            # 35 × 35 × 32
            tower_conv2_2 = slim.conv2d(tower_conv2_1, 32, 3, scope='Conv2d_0c_3x3')
        # 35 × 35 × 96
        mixed = tf.concat([tower_conv, tower_conv1_1, tower_conv2_2], 3)
        # 35 × 35 × 256
        up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,activation_fn=None, scope='Conv2d_1x1')
        # 使用殘差網路scale = 0.17
        net += scale * up
        if 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

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21
• 22
• 23
• 24
• 25
• 26
• 27
• 28
• 29
• 30
• 31
• 32
• 33

Reduction-A結構

　　Reduction-A中含有4個引數k、l、 m、 n，它們對應的值分別為：192, 192, 256, 384，在該層網路結構，輸入為35×35×256，輸出為17×17×896.

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

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21
• 22
• 23
• 24
• 25
• 26
• 27
• 28

Inception-Resnet-B

　　Inception-Resnet-B模組是要重複10次，輸入為17×17×896，輸出為17×17×896，網路結構為：

# 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*128
            tower_conv = slim.conv2d(net, 128, 1, scope='Conv2d_1x1')
        with tf.variable_scope('Branch_1'):
            # 17*17*128
            tower_conv1_0 = slim.conv2d(net, 128, 1, scope='Conv2d_0a_1x1')
            # 17*17*128
            tower_conv1_1 = slim.conv2d(tower_conv1_0, 128, [1, 7],
                                        scope='Conv2d_0b_1x7')
            # 17*17*128
            tower_conv1_2 = slim.conv2d(tower_conv1_1, 128, [7, 1],
                                        scope='Conv2d_0c_7x1')
        # 17*17*256
        mixed = tf.concat([tower_conv, tower_conv1_2], 3)
        # 17*17*896
        up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,activation_fn=None, scope='Conv2d_1x1')
        net += scale * up
        if 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

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21
• 22
• 23
• 24
• 25
• 26
• 27
• 28

Reduction-B

　　Reduction-B的輸入為17*17*896，輸出為8*8*1792。網路結構為：

對應的程式碼為：

def reduction_b(net):
    with tf.variable_scope('Branch_0'):
        # 17*17*256
        tower_conv = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        # 8*8*384
        tower_conv_1 = slim.conv2d(tower_conv, 384, 3, stride=2,
                                   padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_1'):
        # 17*17*256
        tower_conv1 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        # 8*8*256
        tower_conv1_1 = slim.conv2d(tower_conv1, 256, 3, stride=2,
                                    padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_2'):
        # 17*17*256
        tower_conv2 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        # 17*17*256
        tower_conv2_1 = slim.conv2d(tower_conv2, 256, 3,
                                    scope='Conv2d_0b_3x3')
        # 8*8*256
        tower_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*896
        tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',
                                     scope='MaxPool_1a_3x3')
    # 8*8*1792
    net = 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

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21
• 22
• 23
• 24
• 25
• 26
• 27
• 28
• 29
• 30
• 31
• 32
• 33
• 34

Inception-Resnet-C結構

　　Inception-Resnet-C結構重複5次。它輸入為8*8*1792，輸出為8*8*1792。對應的結構為：

　　對應的程式碼為：

# 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*192
            tower_conv = slim.conv2d(net, 192, 1, scope='Conv2d_1x1')
        with tf.variable_scope('Branch_1'):
            # 8*8*192
            tower_conv1_0 = slim.conv2d(net, 192, 1, scope='Conv2d_0a_1x1')
            # 8*8*192
            tower_conv1_1 = slim.conv2d(tower_conv1_0, 192, [1, 3],
                                        scope='Conv2d_0b_1x3')
            # 8*8*192
            tower_conv1_2 = slim.conv2d(tower_conv1_1, 192, [3, 1],
                                        scope='Conv2d_0c_3x1')
        # 8*8*384
        mixed = tf.concat([tower_conv, tower_conv1_2], 3)
        # 8*8*1792
        up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,activation_fn=None, scope='Conv2d_1x1')
        # scale=0.20
        net += scale * up
        if 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

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21
• 22
• 23
• 24
• 25
• 26
• 27
• 28

　　但是在facenet中，接下來又是一層Inception-Resnet-C，但是它沒有重複，並且沒有啟用函式。輸入與輸出大小相同。

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

• 1
• 2

結果輸出

　　結果輸出包含Average Pooling和Dropout (keep 0.8)及Softmax三層，這裡我們以facenet中為例：具體的程式碼如下：

with tf.variable_scope('Logits'):
     end_points['PrePool'] = net
     #pylint: disable=no-member
     # Average Pooling層，輸出為8×8×1792
     net = slim.avg_pool2d(net, net.get_shape()[1:3], padding='VALID',scope='AvgPool_1a_8x8')
     #扁平除了batch_size維度的其它維度。使輸出變為：[batch_size, ...]
     net = slim.flatten(net)
     #dropout層
     net = slim.dropout(net, dropout_keep_prob, is_training=is_training,scope='Dropout')
     end_points['PreLogitsFlatten'] = net
     # 全連結層。輸出為batch_size×128
     net = slim.fully_connected(net, bottleneck_layer_size, activation_fn=None,scope='Bottleneck', reuse=False)