参考论文：Identity Mappings in Deep Residual Networks作者：==Kaiming He==, Xiangyu Zhang, Shaoqing Ren, Jian Sun
这篇是对ResNet经典之作Deep Residual Learning for Image Recognition的一些改进。
何恺明大神在这篇论文中提出了一种新的残差单元。我们将这篇论文中的ResNet结构称为ResNet-V2。
1、残差单元优化
灰色箭头表示信息最容易传播的路径。右:1001-layerResNets的CIFAR-10训练曲线。实线表示测试误差(右边的y轴)，虚线表示训练损失(左边的y轴)。新的残差单元使ResNet-1001更容易训练。
图中的 Iterations 表示迭代次数；Test Error 表示测试集错误率。
(a)original 表示原始的 ResNet 的残差结构，(b)proposed 表示新的 ResNet 的残差结构。主要差别就是(a)结构先卷积后进行 BN 和激活函数计算，最后执行 addition 后再进行ReLU 计算； ==(b)结构先进行 BN 和激活函数计算后卷积，把 addition 后的 ReLU 计算放到了残差结构内部。==
作者使用这两种不同的结构在 CIFAR-10 数据集上做测试，模型用的是 1001层的 ResNet 模型。从图中结果我们可以看出，(b)proposed 的测试集错误率明显更低一些，达到了 4.92%的错误率，(a)original 的测试集错误率是 7.61%。
2、关于残差结构的不同尝试
图2。表1中使用的各种类型的快捷连接。灰色箭头表示信息传播的最容易路径。(b-f)中的快捷连接被不同的组件阻碍。为了简化插图，我们不显示BN层，这里所有单位均采用权值层之后的BN层。
图中(a),(b),(c),(d),(e),(f)都是作者对残差结构的 shortcut 部分进行的不同尝试 ，作者对不同 shortcut 结构的尝试结果如下表所示 。
表1。使用ResNet-110[1]在CIFAR-10测试集上的分类错误，对所有残差单元应用了不同类型的shortcut connections。当测试误差高于20%时，我们报告“失败”。
作者用不同 shortcut 结构的 ResNet-110 在 CIFAR-10 数据集上做测试，发现最原始的(a)original 结构是最好的，也就是 identity mapping 恒等映射是最好的
3、关于激活的尝试
图4。激活的各种用法见表2。所有这些单元都由相同的组件组成，只是顺序不同
表2。使用不同激活函数的CIFAR-10测试集上的分类误差(%)。
最好的结果是(e)full pre-activation，其次到(a)original。(a)original 的残差结构是应用在最原始的 ResNet 中的残差结构；(e)full pre-activation 的残差结构就是我们前面介绍的ResNet-V2 中的残差结构。
4、模型复现
这里我是仿照的Tensorflow官方源代码实现的，个人对代码做了一些精简，因为源码中考虑到了很多情况，这里就简单复现下。注意：ResNet50V2、ResNet101V2与ResNet152V2的搭建方式完全一样，区别在于堆叠Residual Block的数量不同。
import tensorflow as tf
import tensorflow.keras.layers as layers
from tensorflow.keras.models import Model4.1 Residual Block
"""
残差块
Arguments:
x: 输入张量
filters: integer, filters of the bottleneck layer.
kernel_size: 默认是3, kernel size of the bottleneck layer.
stride: default 1, stride of the first layer.
conv_shortcut: default False, use convolution shortcut if True,
otherwise identity shortcut.
name: string, block label.
Returns:
Output tensor for the residual block.
"""
def block2(x, filters, kernel_size=3, stride=1, conv_shortcut=False, name=None):
preact = layers.BatchNormalization(name=name + '_preact_bn')(x)
preact = layers.Activation('relu', name=name + '_preact_relu')(preact)
if conv_shortcut:
shortcut = layers.Conv2D(4 * filters, 1, strides=stride, name=name + '_0_conv')(preact)
else:
shortcut = layers.MaxPooling2D(1, strides=stride)(x) if stride > 1 else x
x = layers.Conv2D(filters, 1, strides=1, use_bias=False, name=name + '_1_conv')(preact)
x = layers.BatchNormalization(name=name + '_1_bn')(x)
x = layers.Activation('relu', name=name + '_1_relu')(x)
x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name=name + '_2_pad')(x)
x = layers.Conv2D(filters,
kernel_size,
strides=stride,
use_bias=False,
name=name + '_2_conv')(x)
x = layers.BatchNormalization(name=name + '_2_bn')(x)
x = layers.Activation('relu', name=name + '_2_relu')(x)
x = layers.Conv2D(4 * filters, 1, name=name + '_3_conv')(x)
x = layers.Add(name=name + '_out')([shortcut, x])
return x我大概对比了下和ResNet50的源代码，发现关于identity shortcut这里稍微有点区别。
ResNet50中是这样的：
if conv_shortcut: shortcut = layers.Conv2D( 4 * filters, 1, strides=stride, name=name + '_0_conv')(x) shortcut = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_0_bn')(shortcut) else: shortcut = x而ResNet50V2中是这样的：
if conv_shortcut: shortcut = layers.Conv2D( 4 * filters, 1, strides=stride, name=name + '_0_conv')(preact) else: shortcut = layers.MaxPooling2D(1, strides=stride)(x) if stride > 1 else x、除此之外，在ResNet50源代码中是没有预激活的(上面代码中的preact)。区别很容易发现，但我还是感觉原版ResNet50比ResNet50V2的代码看起来更优雅。
4.2 堆叠Residual Block
def stack2(x, filters, blocks, stride1=2, name=None):
x = block2(x, filters, conv_shortcut=True, name=name + '_block1')
for i in range(2, blocks):
x = block2(x, filters, name=name + '_block' + str(i))
x = block2(x, filters, stride=stride1, name=name + '_block' + str(blocks))
return x4.3 ResNet50V2架构复现
def ResNet50V2(include_top=True, # 是否包含位于网络顶部的全连接层
preact=True, # 是否使用预激活
use_bias=True, # 是否对卷积层使用偏置
weights='imagenet',
input_tensor=None, # 可选的keras张量，用作模型的图像输入
input_shape=None,
pooling=None,
classes=1000, # 用于分类图像的可选类数
classifier_activation='softmax'): # 分类层激活函数
img_input = layers.Input(shape=input_shape)
x = layers.ZeroPadding2D(padding=((3, 3), (3, 3)), name='conv1_pad')(img_input)
x = layers.Conv2D(64, 7, strides=2, use_bias=use_bias, name='conv1_conv')(x)
if not preact:
x = layers.BatchNormalization(name='conv1_bn')(x)
x = layers.Activation('relu', name='conv1_relu')(x)
x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name='pool1_pad')(x)
x = layers.MaxPooling2D(3, strides=2, name='pool1_pool')(x)
x = stack2(x, 64, 3, name='conv2')
x = stack2(x, 128, 4, name='conv3')
x = stack2(x, 256, 6, name='conv4')
x = stack2(x, 512, 3, stride1=1, name='conv5')
if preact:
x = layers.BatchNormalization(name='post_bn')(x)
x = layers.Activation('relu', name='post_relu')(x)
if include_top:
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D(name='max_pool')(x)
model = Model(img_input, x)
return modelif __name__ == '__main__':
model = ResNet50V2(input_shape=(224, 224, 3))
model.summary()Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 224, 224, 3) 0
__________________________________________________________________________________________________
conv1_pad (ZeroPadding2D) (None, 230, 230, 3) 0 input_1[0][0]
__________________________________________________________________________________________________
conv1_conv (Conv2D) (None, 112, 112, 64) 9472 conv1_pad[0][0]
__________________________________________________________________________________________________
pool1_pad (ZeroPadding2D) (None, 114, 114, 64) 0 conv1_conv[0][0]
__________________________________________________________________________________________________
pool1_pool (MaxPooling2D) (None, 56, 56, 64) 0 pool1_pad[0][0]
__________________________________________________________________________________________________
conv2_block1_preact_bn (BatchNo (None, 56, 56, 64) 256 pool1_pool[0][0]
__________________________________________________________________________________________________
conv2_block1_preact_relu (Activ (None, 56, 56, 64) 0 conv2_block1_preact_bn[0][0]
__________________________________________________________________________________________________
conv2_block1_1_conv (Conv2D) (None, 56, 56, 64) 4096 conv2_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv2_block1_1_bn (BatchNormali (None, 56, 56, 64) 256 conv2_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block1_1_relu (Activation (None, 56, 56, 64) 0 conv2_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block1_2_pad (ZeroPadding (None, 58, 58, 64) 0 conv2_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block1_2_conv (Conv2D) (None, 56, 56, 64) 36864 conv2_block1_2_pad[0][0]
__________________________________________________________________________________________________
conv2_block1_2_bn (BatchNormali (None, 56, 56, 64) 256 conv2_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block1_2_relu (Activation (None, 56, 56, 64) 0 conv2_block1_2_bn[0][0]
__________________________________________________________________________________________________
conv2_block1_0_conv (Conv2D) (None, 56, 56, 256) 16640 conv2_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv2_block1_3_conv (Conv2D) (None, 56, 56, 256) 16640 conv2_block1_2_relu[0][0]
__________________________________________________________________________________________________
conv2_block1_out (Add) (None, 56, 56, 256) 0 conv2_block1_0_conv[0][0]
conv2_block1_3_conv[0][0]
__________________________________________________________________________________________________
conv2_block2_preact_bn (BatchNo (None, 56, 56, 256) 1024 conv2_block1_out[0][0]
__________________________________________________________________________________________________
conv2_block2_preact_relu (Activ (None, 56, 56, 256) 0 conv2_block2_preact_bn[0][0]
__________________________________________________________________________________________________
conv2_block2_1_conv (Conv2D) (None, 56, 56, 64) 16384 conv2_block2_preact_relu[0][0]
__________________________________________________________________________________________________
conv2_block2_1_bn (BatchNormali (None, 56, 56, 64) 256 conv2_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block2_1_relu (Activation (None, 56, 56, 64) 0 conv2_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block2_2_pad (ZeroPadding (None, 58, 58, 64) 0 conv2_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block2_2_conv (Conv2D) (None, 56, 56, 64) 36864 conv2_block2_2_pad[0][0]
__________________________________________________________________________________________________
conv2_block2_2_bn (BatchNormali (None, 56, 56, 64) 256 conv2_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block2_2_relu (Activation (None, 56, 56, 64) 0 conv2_block2_2_bn[0][0]
__________________________________________________________________________________________________
conv2_block2_3_conv (Conv2D) (None, 56, 56, 256) 16640 conv2_block2_2_relu[0][0]
__________________________________________________________________________________________________
conv2_block2_out (Add) (None, 56, 56, 256) 0 conv2_block1_out[0][0]
conv2_block2_3_conv[0][0]
__________________________________________________________________________________________________
conv2_block3_preact_bn (BatchNo (None, 56, 56, 256) 1024 conv2_block2_out[0][0]
__________________________________________________________________________________________________
conv2_block3_preact_relu (Activ (None, 56, 56, 256) 0 conv2_block3_preact_bn[0][0]
__________________________________________________________________________________________________
conv2_block3_1_conv (Conv2D) (None, 56, 56, 64) 16384 conv2_block3_preact_relu[0][0]
__________________________________________________________________________________________________
conv2_block3_1_bn (BatchNormali (None, 56, 56, 64) 256 conv2_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv2_block3_1_relu (Activation (None, 56, 56, 64) 0 conv2_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv2_block3_2_pad (ZeroPadding (None, 58, 58, 64) 0 conv2_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv2_block3_2_conv (Conv2D) (None, 28, 28, 64) 36864 conv2_block3_2_pad[0][0]
__________________________________________________________________________________________________
conv2_block3_2_bn (BatchNormali (None, 28, 28, 64) 256 conv2_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv2_block3_2_relu (Activation (None, 28, 28, 64) 0 conv2_block3_2_bn[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 28, 28, 256) 0 conv2_block2_out[0][0]
__________________________________________________________________________________________________
conv2_block3_3_conv (Conv2D) (None, 28, 28, 256) 16640 conv2_block3_2_relu[0][0]
__________________________________________________________________________________________________
conv2_block3_out (Add) (None, 28, 28, 256) 0 max_pooling2d[0][0]
conv2_block3_3_conv[0][0]
__________________________________________________________________________________________________
conv3_block1_preact_bn (BatchNo (None, 28, 28, 256) 1024 conv2_block3_out[0][0]
__________________________________________________________________________________________________
conv3_block1_preact_relu (Activ (None, 28, 28, 256) 0 conv3_block1_preact_bn[0][0]
__________________________________________________________________________________________________
conv3_block1_1_conv (Conv2D) (None, 28, 28, 128) 32768 conv3_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv3_block1_1_bn (BatchNormali (None, 28, 28, 128) 512 conv3_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block1_1_relu (Activation (None, 28, 28, 128) 0 conv3_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block1_2_pad (ZeroPadding (None, 30, 30, 128) 0 conv3_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block1_2_conv (Conv2D) (None, 28, 28, 128) 147456 conv3_block1_2_pad[0][0]
__________________________________________________________________________________________________
conv3_block1_2_bn (BatchNormali (None, 28, 28, 128) 512 conv3_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block1_2_relu (Activation (None, 28, 28, 128) 0 conv3_block1_2_bn[0][0]
__________________________________________________________________________________________________
conv3_block1_0_conv (Conv2D) (None, 28, 28, 512) 131584 conv3_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv3_block1_3_conv (Conv2D) (None, 28, 28, 512) 66048 conv3_block1_2_relu[0][0]
__________________________________________________________________________________________________
conv3_block1_out (Add) (None, 28, 28, 512) 0 conv3_block1_0_conv[0][0]
conv3_block1_3_conv[0][0]
__________________________________________________________________________________________________
conv3_block2_preact_bn (BatchNo (None, 28, 28, 512) 2048 conv3_block1_out[0][0]
__________________________________________________________________________________________________
conv3_block2_preact_relu (Activ (None, 28, 28, 512) 0 conv3_block2_preact_bn[0][0]
__________________________________________________________________________________________________
conv3_block2_1_conv (Conv2D) (None, 28, 28, 128) 65536 conv3_block2_preact_relu[0][0]
__________________________________________________________________________________________________
conv3_block2_1_bn (BatchNormali (None, 28, 28, 128) 512 conv3_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block2_1_relu (Activation (None, 28, 28, 128) 0 conv3_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block2_2_pad (ZeroPadding (None, 30, 30, 128) 0 conv3_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block2_2_conv (Conv2D) (None, 28, 28, 128) 147456 conv3_block2_2_pad[0][0]
__________________________________________________________________________________________________
conv3_block2_2_bn (BatchNormali (None, 28, 28, 128) 512 conv3_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block2_2_relu (Activation (None, 28, 28, 128) 0 conv3_block2_2_bn[0][0]
__________________________________________________________________________________________________
conv3_block2_3_conv (Conv2D) (None, 28, 28, 512) 66048 conv3_block2_2_relu[0][0]
__________________________________________________________________________________________________
conv3_block2_out (Add) (None, 28, 28, 512) 0 conv3_block1_out[0][0]
conv3_block2_3_conv[0][0]
__________________________________________________________________________________________________
conv3_block3_preact_bn (BatchNo (None, 28, 28, 512) 2048 conv3_block2_out[0][0]
__________________________________________________________________________________________________
conv3_block3_preact_relu (Activ (None, 28, 28, 512) 0 conv3_block3_preact_bn[0][0]
__________________________________________________________________________________________________
conv3_block3_1_conv (Conv2D) (None, 28, 28, 128) 65536 conv3_block3_preact_relu[0][0]
__________________________________________________________________________________________________
conv3_block3_1_bn (BatchNormali (None, 28, 28, 128) 512 conv3_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block3_1_relu (Activation (None, 28, 28, 128) 0 conv3_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block3_2_pad (ZeroPadding (None, 30, 30, 128) 0 conv3_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block3_2_conv (Conv2D) (None, 28, 28, 128) 147456 conv3_block3_2_pad[0][0]
__________________________________________________________________________________________________
conv3_block3_2_bn (BatchNormali (None, 28, 28, 128) 512 conv3_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block3_2_relu (Activation (None, 28, 28, 128) 0 conv3_block3_2_bn[0][0]
__________________________________________________________________________________________________
conv3_block3_3_conv (Conv2D) (None, 28, 28, 512) 66048 conv3_block3_2_relu[0][0]
__________________________________________________________________________________________________
conv3_block3_out (Add) (None, 28, 28, 512) 0 conv3_block2_out[0][0]
conv3_block3_3_conv[0][0]
__________________________________________________________________________________________________
conv3_block4_preact_bn (BatchNo (None, 28, 28, 512) 2048 conv3_block3_out[0][0]
__________________________________________________________________________________________________
conv3_block4_preact_relu (Activ (None, 28, 28, 512) 0 conv3_block4_preact_bn[0][0]
__________________________________________________________________________________________________
conv3_block4_1_conv (Conv2D) (None, 28, 28, 128) 65536 conv3_block4_preact_relu[0][0]
__________________________________________________________________________________________________
conv3_block4_1_bn (BatchNormali (None, 28, 28, 128) 512 conv3_block4_1_conv[0][0]
__________________________________________________________________________________________________
conv3_block4_1_relu (Activation (None, 28, 28, 128) 0 conv3_block4_1_bn[0][0]
__________________________________________________________________________________________________
conv3_block4_2_pad (ZeroPadding (None, 30, 30, 128) 0 conv3_block4_1_relu[0][0]
__________________________________________________________________________________________________
conv3_block4_2_conv (Conv2D) (None, 14, 14, 128) 147456 conv3_block4_2_pad[0][0]
__________________________________________________________________________________________________
conv3_block4_2_bn (BatchNormali (None, 14, 14, 128) 512 conv3_block4_2_conv[0][0]
__________________________________________________________________________________________________
conv3_block4_2_relu (Activation (None, 14, 14, 128) 0 conv3_block4_2_bn[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 14, 14, 512) 0 conv3_block3_out[0][0]
__________________________________________________________________________________________________
conv3_block4_3_conv (Conv2D) (None, 14, 14, 512) 66048 conv3_block4_2_relu[0][0]
__________________________________________________________________________________________________
conv3_block4_out (Add) (None, 14, 14, 512) 0 max_pooling2d_1[0][0]
conv3_block4_3_conv[0][0]
__________________________________________________________________________________________________
conv4_block1_preact_bn (BatchNo (None, 14, 14, 512) 2048 conv3_block4_out[0][0]
__________________________________________________________________________________________________
conv4_block1_preact_relu (Activ (None, 14, 14, 512) 0 conv4_block1_preact_bn[0][0]
__________________________________________________________________________________________________
conv4_block1_1_conv (Conv2D) (None, 14, 14, 256) 131072 conv4_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv4_block1_1_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block1_1_relu (Activation (None, 14, 14, 256) 0 conv4_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block1_2_pad (ZeroPadding (None, 16, 16, 256) 0 conv4_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block1_2_conv (Conv2D) (None, 14, 14, 256) 589824 conv4_block1_2_pad[0][0]
__________________________________________________________________________________________________
conv4_block1_2_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block1_2_relu (Activation (None, 14, 14, 256) 0 conv4_block1_2_bn[0][0]
__________________________________________________________________________________________________
conv4_block1_0_conv (Conv2D) (None, 14, 14, 1024) 525312 conv4_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv4_block1_3_conv (Conv2D) (None, 14, 14, 1024) 263168 conv4_block1_2_relu[0][0]
__________________________________________________________________________________________________
conv4_block1_out (Add) (None, 14, 14, 1024) 0 conv4_block1_0_conv[0][0]
conv4_block1_3_conv[0][0]
__________________________________________________________________________________________________
conv4_block2_preact_bn (BatchNo (None, 14, 14, 1024) 4096 conv4_block1_out[0][0]
__________________________________________________________________________________________________
conv4_block2_preact_relu (Activ (None, 14, 14, 1024) 0 conv4_block2_preact_bn[0][0]
__________________________________________________________________________________________________
conv4_block2_1_conv (Conv2D) (None, 14, 14, 256) 262144 conv4_block2_preact_relu[0][0]
__________________________________________________________________________________________________
conv4_block2_1_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block2_1_relu (Activation (None, 14, 14, 256) 0 conv4_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block2_2_pad (ZeroPadding (None, 16, 16, 256) 0 conv4_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block2_2_conv (Conv2D) (None, 14, 14, 256) 589824 conv4_block2_2_pad[0][0]
__________________________________________________________________________________________________
conv4_block2_2_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block2_2_relu (Activation (None, 14, 14, 256) 0 conv4_block2_2_bn[0][0]
__________________________________________________________________________________________________
conv4_block2_3_conv (Conv2D) (None, 14, 14, 1024) 263168 conv4_block2_2_relu[0][0]
__________________________________________________________________________________________________
conv4_block2_out (Add) (None, 14, 14, 1024) 0 conv4_block1_out[0][0]
conv4_block2_3_conv[0][0]
__________________________________________________________________________________________________
conv4_block3_preact_bn (BatchNo (None, 14, 14, 1024) 4096 conv4_block2_out[0][0]
__________________________________________________________________________________________________
conv4_block3_preact_relu (Activ (None, 14, 14, 1024) 0 conv4_block3_preact_bn[0][0]
__________________________________________________________________________________________________
conv4_block3_1_conv (Conv2D) (None, 14, 14, 256) 262144 conv4_block3_preact_relu[0][0]
__________________________________________________________________________________________________
conv4_block3_1_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block3_1_relu (Activation (None, 14, 14, 256) 0 conv4_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block3_2_pad (ZeroPadding (None, 16, 16, 256) 0 conv4_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block3_2_conv (Conv2D) (None, 14, 14, 256) 589824 conv4_block3_2_pad[0][0]
__________________________________________________________________________________________________
conv4_block3_2_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block3_2_relu (Activation (None, 14, 14, 256) 0 conv4_block3_2_bn[0][0]
__________________________________________________________________________________________________
conv4_block3_3_conv (Conv2D) (None, 14, 14, 1024) 263168 conv4_block3_2_relu[0][0]
__________________________________________________________________________________________________
conv4_block3_out (Add) (None, 14, 14, 1024) 0 conv4_block2_out[0][0]
conv4_block3_3_conv[0][0]
__________________________________________________________________________________________________
conv4_block4_preact_bn (BatchNo (None, 14, 14, 1024) 4096 conv4_block3_out[0][0]
__________________________________________________________________________________________________
conv4_block4_preact_relu (Activ (None, 14, 14, 1024) 0 conv4_block4_preact_bn[0][0]
__________________________________________________________________________________________________
conv4_block4_1_conv (Conv2D) (None, 14, 14, 256) 262144 conv4_block4_preact_relu[0][0]
__________________________________________________________________________________________________
conv4_block4_1_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block4_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block4_1_relu (Activation (None, 14, 14, 256) 0 conv4_block4_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block4_2_pad (ZeroPadding (None, 16, 16, 256) 0 conv4_block4_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block4_2_conv (Conv2D) (None, 14, 14, 256) 589824 conv4_block4_2_pad[0][0]
__________________________________________________________________________________________________
conv4_block4_2_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block4_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block4_2_relu (Activation (None, 14, 14, 256) 0 conv4_block4_2_bn[0][0]
__________________________________________________________________________________________________
conv4_block4_3_conv (Conv2D) (None, 14, 14, 1024) 263168 conv4_block4_2_relu[0][0]
__________________________________________________________________________________________________
conv4_block4_out (Add) (None, 14, 14, 1024) 0 conv4_block3_out[0][0]
conv4_block4_3_conv[0][0]
__________________________________________________________________________________________________
conv4_block5_preact_bn (BatchNo (None, 14, 14, 1024) 4096 conv4_block4_out[0][0]
__________________________________________________________________________________________________
conv4_block5_preact_relu (Activ (None, 14, 14, 1024) 0 conv4_block5_preact_bn[0][0]
__________________________________________________________________________________________________
conv4_block5_1_conv (Conv2D) (None, 14, 14, 256) 262144 conv4_block5_preact_relu[0][0]
__________________________________________________________________________________________________
conv4_block5_1_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block5_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block5_1_relu (Activation (None, 14, 14, 256) 0 conv4_block5_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block5_2_pad (ZeroPadding (None, 16, 16, 256) 0 conv4_block5_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block5_2_conv (Conv2D) (None, 14, 14, 256) 589824 conv4_block5_2_pad[0][0]
__________________________________________________________________________________________________
conv4_block5_2_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block5_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block5_2_relu (Activation (None, 14, 14, 256) 0 conv4_block5_2_bn[0][0]
__________________________________________________________________________________________________
conv4_block5_3_conv (Conv2D) (None, 14, 14, 1024) 263168 conv4_block5_2_relu[0][0]
__________________________________________________________________________________________________
conv4_block5_out (Add) (None, 14, 14, 1024) 0 conv4_block4_out[0][0]
conv4_block5_3_conv[0][0]
__________________________________________________________________________________________________
conv4_block6_preact_bn (BatchNo (None, 14, 14, 1024) 4096 conv4_block5_out[0][0]
__________________________________________________________________________________________________
conv4_block6_preact_relu (Activ (None, 14, 14, 1024) 0 conv4_block6_preact_bn[0][0]
__________________________________________________________________________________________________
conv4_block6_1_conv (Conv2D) (None, 14, 14, 256) 262144 conv4_block6_preact_relu[0][0]
__________________________________________________________________________________________________
conv4_block6_1_bn (BatchNormali (None, 14, 14, 256) 1024 conv4_block6_1_conv[0][0]
__________________________________________________________________________________________________
conv4_block6_1_relu (Activation (None, 14, 14, 256) 0 conv4_block6_1_bn[0][0]
__________________________________________________________________________________________________
conv4_block6_2_pad (ZeroPadding (None, 16, 16, 256) 0 conv4_block6_1_relu[0][0]
__________________________________________________________________________________________________
conv4_block6_2_conv (Conv2D) (None, 7, 7, 256) 589824 conv4_block6_2_pad[0][0]
__________________________________________________________________________________________________
conv4_block6_2_bn (BatchNormali (None, 7, 7, 256) 1024 conv4_block6_2_conv[0][0]
__________________________________________________________________________________________________
conv4_block6_2_relu (Activation (None, 7, 7, 256) 0 conv4_block6_2_bn[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 7, 7, 1024) 0 conv4_block5_out[0][0]
__________________________________________________________________________________________________
conv4_block6_3_conv (Conv2D) (None, 7, 7, 1024) 263168 conv4_block6_2_relu[0][0]
__________________________________________________________________________________________________
conv4_block6_out (Add) (None, 7, 7, 1024) 0 max_pooling2d_2[0][0]
conv4_block6_3_conv[0][0]
__________________________________________________________________________________________________
conv5_block1_preact_bn (BatchNo (None, 7, 7, 1024) 4096 conv4_block6_out[0][0]
__________________________________________________________________________________________________
conv5_block1_preact_relu (Activ (None, 7, 7, 1024) 0 conv5_block1_preact_bn[0][0]
__________________________________________________________________________________________________
conv5_block1_1_conv (Conv2D) (None, 7, 7, 512) 524288 conv5_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv5_block1_1_bn (BatchNormali (None, 7, 7, 512) 2048 conv5_block1_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block1_1_relu (Activation (None, 7, 7, 512) 0 conv5_block1_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block1_2_pad (ZeroPadding (None, 9, 9, 512) 0 conv5_block1_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block1_2_conv (Conv2D) (None, 7, 7, 512) 2359296 conv5_block1_2_pad[0][0]
__________________________________________________________________________________________________
conv5_block1_2_bn (BatchNormali (None, 7, 7, 512) 2048 conv5_block1_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block1_2_relu (Activation (None, 7, 7, 512) 0 conv5_block1_2_bn[0][0]
__________________________________________________________________________________________________
conv5_block1_0_conv (Conv2D) (None, 7, 7, 2048) 2099200 conv5_block1_preact_relu[0][0]
__________________________________________________________________________________________________
conv5_block1_3_conv (Conv2D) (None, 7, 7, 2048) 1050624 conv5_block1_2_relu[0][0]
__________________________________________________________________________________________________
conv5_block1_out (Add) (None, 7, 7, 2048) 0 conv5_block1_0_conv[0][0]
conv5_block1_3_conv[0][0]
__________________________________________________________________________________________________
conv5_block2_preact_bn (BatchNo (None, 7, 7, 2048) 8192 conv5_block1_out[0][0]
__________________________________________________________________________________________________
conv5_block2_preact_relu (Activ (None, 7, 7, 2048) 0 conv5_block2_preact_bn[0][0]
__________________________________________________________________________________________________
conv5_block2_1_conv (Conv2D) (None, 7, 7, 512) 1048576 conv5_block2_preact_relu[0][0]
__________________________________________________________________________________________________
conv5_block2_1_bn (BatchNormali (None, 7, 7, 512) 2048 conv5_block2_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block2_1_relu (Activation (None, 7, 7, 512) 0 conv5_block2_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block2_2_pad (ZeroPadding (None, 9, 9, 512) 0 conv5_block2_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block2_2_conv (Conv2D) (None, 7, 7, 512) 2359296 conv5_block2_2_pad[0][0]
__________________________________________________________________________________________________
conv5_block2_2_bn (BatchNormali (None, 7, 7, 512) 2048 conv5_block2_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block2_2_relu (Activation (None, 7, 7, 512) 0 conv5_block2_2_bn[0][0]
__________________________________________________________________________________________________
conv5_block2_3_conv (Conv2D) (None, 7, 7, 2048) 1050624 conv5_block2_2_relu[0][0]
__________________________________________________________________________________________________
conv5_block2_out (Add) (None, 7, 7, 2048) 0 conv5_block1_out[0][0]
conv5_block2_3_conv[0][0]
__________________________________________________________________________________________________
conv5_block3_preact_bn (BatchNo (None, 7, 7, 2048) 8192 conv5_block2_out[0][0]
__________________________________________________________________________________________________
conv5_block3_preact_relu (Activ (None, 7, 7, 2048) 0 conv5_block3_preact_bn[0][0]
__________________________________________________________________________________________________
conv5_block3_1_conv (Conv2D) (None, 7, 7, 512) 1048576 conv5_block3_preact_relu[0][0]
__________________________________________________________________________________________________
conv5_block3_1_bn (BatchNormali (None, 7, 7, 512) 2048 conv5_block3_1_conv[0][0]
__________________________________________________________________________________________________
conv5_block3_1_relu (Activation (None, 7, 7, 512) 0 conv5_block3_1_bn[0][0]
__________________________________________________________________________________________________
conv5_block3_2_pad (ZeroPadding (None, 9, 9, 512) 0 conv5_block3_1_relu[0][0]
__________________________________________________________________________________________________
conv5_block3_2_conv (Conv2D) (None, 7, 7, 512) 2359296 conv5_block3_2_pad[0][0]
__________________________________________________________________________________________________
conv5_block3_2_bn (BatchNormali (None, 7, 7, 512) 2048 conv5_block3_2_conv[0][0]
__________________________________________________________________________________________________
conv5_block3_2_relu (Activation (None, 7, 7, 512) 0 conv5_block3_2_bn[0][0]
__________________________________________________________________________________________________
conv5_block3_3_conv (Conv2D) (None, 7, 7, 2048) 1050624 conv5_block3_2_relu[0][0]
__________________________________________________________________________________________________
conv5_block3_out (Add) (None, 7, 7, 2048) 0 conv5_block2_out[0][0]
conv5_block3_3_conv[0][0]
__________________________________________________________________________________________________
post_bn (BatchNormalization) (None, 7, 7, 2048) 8192 conv5_block3_out[0][0]
__________________________________________________________________________________________________
post_relu (Activation) (None, 7, 7, 2048) 0 post_bn[0][0]
__________________________________________________________________________________________________
avg_pool (GlobalAveragePooling2 (None, 2048) 0 post_relu[0][0]
__________________________________________________________________________________________________
predictions (Dense) (None, 1000) 2049000 avg_pool[0][0]
==================================================================================================
Total params: 25,613,800
Trainable params: 25,568,360
Non-trainable params: 45,4404.4 ResNet101V2架构复现
其实和上面的搭建方式完全一样，没必要再写了，但是为了对比还是写一下。
def ResNet101V2(include_top=True, # 是否包含位于网络顶部的全连接层
preact=True, # 是否使用预激活
use_bias=True, # 是否对卷积层使用偏置
weights='imagenet',
input_tensor=None, # 可选的keras张量，用作模型的图像输入
input_shape=None,
pooling=None,
classes=1000, # 用于分类图像的可选类数
classifier_activation='softmax'): # 分类层激活函数
img_input = layers.Input(shape=input_shape)
x = layers.ZeroPadding2D(padding=((3, 3), (3, 3)), name='conv1_pad')(img_input)
x = layers.Conv2D(64, 7, strides=2, use_bias=use_bias, name='conv1_conv')(x)
if not preact:
x = layers.BatchNormalization(name='conv1_bn')(x)
x = layers.Activation('relu', name='conv1_relu')(x)
x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name='pool1_pad')(x)
x = layers.MaxPooling2D(3, strides=2, name='pool1_pool')(x)
x = stack2(x, 64, 3, name='conv2')
x = stack2(x, 128, 4, name='conv3')
x = stack2(x, 256, 23, name='conv4')
x = stack2(x, 512, 3, stride1=1, name='conv5')
if preact:
x = layers.BatchNormalization(name='post_bn')(x)
x = layers.Activation('relu', name='post_relu')(x)
if include_top:
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D(name='max_pool')(x)
model = Model(img_input, x)
return model4.5 ResNet152V2模型复现
只需要将堆叠快部分的代码改为如下即可。
x = stack2(x, 64, 3, name='conv2')
x = stack2(x, 128, 8, name='conv3')
x = stack2(x, 256, 36, name='conv4')
x = stack2(x, 512, 3, stride1=1, name='conv5')4.6 ResNet50V2模型结构大图
4.7 ResNet101V2模型结构大图
4.8 ResNet152模型结构大图
References
Identity Mappings in Deep Residual Networks
Deep Residual Learning for Image Recognition
https://github.com/tensorflow/tensorflow/blob/b36436b087bd8e8701ef51718179037cccdfc26e/tensorflow/python/keras/applications/resnet.py#L293
