Pytorch从零开始实战16

Pytorch从零开始实战——ResNeXt-50算法的思考

本系列来源于365天深度学习训练营

原作者K同学

对于上次ResNeXt-50算法，我们同样有基于TensorFlow的实现。具体代码如下。

引入头文件

import numpy as npfrom tensorflow.keras.preprocessing.image import ImageDataGeneratorfrom tensorflow.keras.utils import to_categoricalfrom tensorflow.keras.models import Sequentialfrom tensorflow.keras.layers import Input, Dense, Dropout, Conv2D, MaxPool2D, Flatten, GlobalAvgPool2D, concatenate, \BatchNormalization, Activation, Add, ZeroPadding2D, Lambdafrom tensorflow.keras.layers import ReLUfrom tensorflow.keras.optimizers import Adamimport matplotlib.pyplot as pltfrom tensorflow.keras.callbacks import LearningRateSchedulerfrom tensorflow.keras.models import Model

分组卷积模块

# 定义分组卷积def grouped_convolution_block(init_x, strides, groups, g_channels):group_list = []# 分组进行卷积for c in range(groups):# 分组取出数据x = Lambda(lambda x: x[:, :, :, c * g_channels:(c + 1) * g_channels])(init_x)# 分组进行卷积x = Conv2D(filters=g_channels, kernel_size=(3, 3),strides=strides, padding='same', use_bias=False)(x)# 存入listgroup_list.append(x)# 合并list中的数据group_merage = concatenate(group_list, axis=3)x = BatchNormalization(epsilon=1.001e-5)(group_merage)x = ReLU()(x)return x

残差单元

# 定义残差单元def block(x, filters, strides=1, groups=32, conv_shortcut=True):if conv_shortcut:shortcut = Conv2D(filters * 2, kernel_size=(1, 1), strides=strides, padding='same', use_bias=False)(x)# epsilon为BN公式中防止分母为零的值shortcut = BatchNormalization(epsilon=1.001e-5)(shortcut)else:# identity_shortcutshortcut = x# 三层卷积层x = Conv2D(filters=filters, kernel_size=(1, 1), strides=1, padding='same', use_bias=False)(x)x = BatchNormalization(epsilon=1.001e-5)(x)x = ReLU()(x)# 计算每组的通道数g_channels = int(filters / groups)# 进行分组卷积x = grouped_convolution_block(x, strides, groups, g_channels)x = Conv2D(filters=filters * 2, kernel_size=(1, 1), strides=1, padding='same', use_bias=False)(x)x = BatchNormalization(epsilon=1.001e-5)(x)x = Add()([x, shortcut])x = ReLU()(x)return x

堆叠残差单元

# 堆叠残差单元def stack(x, filters, blocks, strides, groups=32):# 每个stack的第一个block的残差连接都需要使用1*1卷积升维x = block(x, filters, strides=strides, groups=groups)for i in range(blocks):x = block(x, filters, groups=groups, conv_shortcut=False)return x

网络搭建

# 定义ResNext50(32*4d)网络def ResNext50(input_shape, num_classes):inputs = Input(shape=input_shape)# 填充3圈0，[224,224,3]->[230,230,3]x = ZeroPadding2D((3, 3))(inputs)x = Conv2D(filters=64, kernel_size=(7, 7), strides=2, padding='valid')(x)x = BatchNormalization(epsilon=1.001e-5)(x)x = ReLU()(x)# 填充1圈0x = ZeroPadding2D((1, 1))(x)x = MaxPool2D(pool_size=(3, 3), strides=2, padding='valid')(x)# 堆叠残差结构x = stack(x, filters=128, blocks=2, strides=1)x = stack(x, filters=256, blocks=3, strides=2)x = stack(x, filters=512, blocks=5, strides=2)x = stack(x, filters=1024, blocks=2, strides=2)# 根据特征图大小进行全局平均池化x = GlobalAvgPool2D()(x)x = Dense(num_classes, activation='softmax')(x)# 定义模型model = Model(inputs=inputs, outputs=x)return model

对于残差单元中的代码，提出一个问题：当conv_shortcut=False的时候，在执行Add操作时，理论上通道数不一致，为什么代码不报错？

答：这主要是跟下面堆叠残差单元的代码有关系，每个stack第一轮总会令conv_shortcut为True，使得x通道数进行扩展，而后面循环的时候传入的filters还是这个函数的实参，没有发生变化，但由于conv_shortcut为False，此时shortcut的通道数是与上面的x一致，所以在Add的时候，代码不会报错。

def stack(x, filters, blocks, strides, groups=32):# 每个stack的第一个block的残差连接都需要使用1*1卷积升维x = block(x, filters, strides=strides, groups=groups)for i in range(blocks):x = block(x, filters, groups=groups, conv_shortcut=False)return x

本文只是对ResNeXt-50算法的部分代码进行思考，学习过程中需要积极思考与探索，以提高能力和解决问题。