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 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
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 × 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')
# 使用残差网络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
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, 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
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*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
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*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
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*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
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8×8×1792
net = 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×128
net = 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~~~~