InceptionNet V3(keras实现)

写在前面：第一次自己看着网络图用代码搭建（记录我的进步哈哈，大佬勿嘲）

----------------------------------------分割线---------------------------------------

一、Google InceptionNet V3结构图

转载一波大佬的结构图，这是大佬文章链接，详细网络模块请移步

二、利用TensorFlow2.x中集成的Keras搭建InceptionNet V3的源代码

import tensorflow as tf
import numpy as np
import cv2 as cv
from PIL import Image
import matplotlib.pyplot as plt
import os
import random as rd
from tensorflow.keras.layers import Input,Conv2D,BatchNormalization,\
    MaxPooling2D,AveragePooling2D,Concatenate,Dense,Flatten,Activation

def build_inception_v3(input_shape=[299,299,3],):
    def my_conv2D(last_layer,my_filters=32,my_kernel_size=(3,3),my_stried=2,my_padding='valid'):
        x = Conv2D(filters=my_filters,kernel_size=my_kernel_size,strides=my_stried,padding=my_padding)(last_layer)
        x = BatchNormalization(scale=False)(x)
        x = Activation('relu')(x)
        return x

    def block1_1(last_layer):
        x1_1 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=64,my_stried=1,my_padding='same')
        x1_2 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=48,my_stried=1,my_padding='same')
        x1_3 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=64,my_stried=1,my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3,3),strides=1,padding='same')(last_layer)

        x2_2 = my_conv2D(x1_2,my_kernel_size=(5,5),my_filters=64,my_stried=1,my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(3,3), my_filters=96, my_stried=1,my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1,1), my_filters=32, my_stried=1,my_padding='same')

        x3_3 = my_conv2D(x2_3, my_kernel_size=(3, 3), my_filters=96, my_stried=1,my_padding='same')

        next_layer = Concatenate()([x1_1,x2_2,x3_3,x2_4])

        return next_layer

    def block1_2(last_layer):
        x1_1 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=64,my_stried=1,my_padding='same')
        x1_2 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=48,my_stried=1,my_padding='same')
        x1_3 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=64,my_stried=1,my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3,3),strides=1,padding='same')(last_layer)

        x2_2 = my_conv2D(x1_2,my_kernel_size=(5,5),my_filters=64,my_stried=1,my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(3,3), my_filters=96, my_stried=1,my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1,1), my_filters=64, my_stried=1,my_padding='same')

        x3_3 = my_conv2D(x2_3, my_kernel_size=(3, 3), my_filters=96, my_stried=1,my_padding='same')

        next_layer = Concatenate()([x1_1,x2_2,x3_3,x2_4])

        return next_layer

    def block1_3(last_layer):
        x1_1 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=64,my_stried=1,my_padding='same')
        x1_2 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=48,my_stried=1,my_padding='same')
        x1_3 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=64,my_stried=1,my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3,3),strides=1,padding='same')(last_layer)

        x2_2 = my_conv2D(x1_2,my_kernel_size=(5,5),my_filters=64,my_stried=1,my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(3,3), my_filters=96, my_stried=1,my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1,1), my_filters=64, my_stried=1,my_padding='same')

        x3_3 = my_conv2D(x2_3, my_kernel_size=(3, 3), my_filters=96, my_stried=1,my_padding='same')

        next_layer = Concatenate()([x1_1,x2_2,x3_3,x2_4])

        return next_layer

    def block2_1(last_layer):
        x1_1 = my_conv2D(last_layer,my_kernel_size=(3,3),my_filters=384,my_stried=2)
        x1_2 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=48,my_stried=1,my_padding='same')
        x1_3 = MaxPooling2D(pool_size=(3,3),strides=2)(last_layer)

        x2_2 = my_conv2D(x1_2,my_kernel_size=(3,3),my_filters=64,my_stried=1,my_padding='same')

        x3_2 = my_conv2D(x2_2, my_kernel_size=(3, 3), my_filters=96, my_stried=2)

        next_layer = Concatenate()([x1_1,x3_2,x1_3])

        return next_layer

    def block2_2(last_layer):
        x1_1 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=192,my_stried=1,my_padding='same')
        x1_2 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=128,my_stried=1,my_padding='same')
        x1_3 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=128,my_stried=1,my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3,3),strides=1,padding='same')(last_layer)

        x2_2 = my_conv2D(x1_2,my_kernel_size=(1,7),my_filters=128,my_stried=1,my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(7, 1), my_filters=128, my_stried=1,my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1, 1), my_filters=192, my_stried=1, my_padding='same')

        x3_2 = my_conv2D(x2_2, my_kernel_size=(7, 1), my_filters=192, my_stried=1, my_padding='same')
        x3_3 = my_conv2D(x2_3, my_kernel_size=(1, 7), my_filters=128, my_stried=1, my_padding='same')

        x4_3 = my_conv2D(x3_3, my_kernel_size=(7, 1), my_filters=128, my_stried=1, my_padding='same')

        x5_3 = my_conv2D(x4_3, my_kernel_size=(1, 7), my_filters=192, my_stried=1, my_padding='same')

        next_layer = Concatenate()([x1_1,x3_2,x5_3,x2_4])

        return next_layer

    def block2_3_4(last_layer):
        x1_1 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=192,my_stried=1,my_padding='same')
        x1_2 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=160,my_stried=1,my_padding='same')
        x1_3 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=160,my_stried=1,my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3,3),strides=1,padding='same')(last_layer)

        x2_2 = my_conv2D(x1_2,my_kernel_size=(1,7),my_filters=160,my_stried=1,my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(7, 1), my_filters=160, my_stried=1,my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1, 1), my_filters=192, my_stried=1, my_padding='same')

        x3_2 = my_conv2D(x2_2, my_kernel_size=(7, 1), my_filters=192, my_stried=1, my_padding='same')
        x3_3 = my_conv2D(x2_3, my_kernel_size=(1, 7), my_filters=160, my_stried=1, my_padding='same')

        x4_3 = my_conv2D(x3_3, my_kernel_size=(7, 1), my_filters=160, my_stried=1, my_padding='same')

        x5_3 = my_conv2D(x4_3, my_kernel_size=(1, 7), my_filters=192, my_stried=1, my_padding='same')

        next_layer = Concatenate()([x1_1,x3_2,x5_3,x2_4])

        return next_layer

    def block2_5(last_layer):
        x1_1 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=192,my_stried=1,my_padding='same')
        x1_2 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=192,my_stried=1,my_padding='same')
        x1_3 = my_conv2D(last_layer,my_kernel_size=(1,1),my_filters=192,my_stried=1,my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3,3),strides=1,padding='same')(last_layer)

        x2_2 = my_conv2D(x1_2,my_kernel_size=(1,7),my_filters=192,my_stried=1,my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(7, 1), my_filters=192, my_stried=1,my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1, 1), my_filters=192, my_stried=1, my_padding='same')

        x3_2 = my_conv2D(x2_2, my_kernel_size=(7, 1), my_filters=192, my_stried=1, my_padding='same')
        x3_3 = my_conv2D(x2_3, my_kernel_size=(1, 7), my_filters=192, my_stried=1, my_padding='same')

        x4_3 = my_conv2D(x3_3, my_kernel_size=(7, 1), my_filters=192, my_stried=1, my_padding='same')

        x5_3 = my_conv2D(x4_3, my_kernel_size=(1, 7), my_filters=192, my_stried=1, my_padding='same')

        next_layer = Concatenate()([x1_1,x3_2,x5_3,x2_4])

        return next_layer

    def block3_1(last_layer):
        x1_1 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=192, my_stried=1, my_padding='same')
        x1_2 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=192, my_stried=1, my_padding='same')
        x1_3 = MaxPooling2D(pool_size=(3,3),strides=2)(last_layer)

        x2_1 = my_conv2D(x1_1, my_kernel_size=(3, 3), my_filters=320, my_stried=2)
        x2_2 = my_conv2D(x1_2, my_kernel_size=(1, 7), my_filters=192, my_stried=1, my_padding='same')

        x3_2 = my_conv2D(x2_2, my_kernel_size=(7, 1), my_filters=192, my_stried=1, my_padding='same')

        x4_2 = my_conv2D(x3_2, my_kernel_size=(3, 3), my_filters=192, my_stried=2)

        next_layer = Concatenate()([x2_1, x4_2, x1_3])

        return next_layer

    def block3_2(last_layer):
        x1_1 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=320, my_stried=1, my_padding='same')
        x1_2 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=320, my_stried=1, my_padding='same')
        x1_3 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=320, my_stried=1, my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3, 3), strides=1, padding='same')(last_layer)

        x2_2_1 = my_conv2D(x1_2, my_kernel_size=(1, 3), my_filters=384, my_stried=1, my_padding='same')
        x2_2_2 = my_conv2D(x1_2, my_kernel_size=(3, 1), my_filters=384, my_stried=1, my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(3, 3), my_filters=384, my_stried=1, my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1, 1), my_filters=192, my_stried=1, my_padding='same')

        x3_2 = Concatenate()([x2_2_1, x2_2_2])
        x3_3_1 = my_conv2D(x2_3, my_kernel_size=(1, 3), my_filters=384, my_stried=1, my_padding='same')
        x3_3_2 = my_conv2D(x2_3, my_kernel_size=(3, 1), my_filters=384, my_stried=1, my_padding='same')

        x4_3 = Concatenate()([x3_3_1, x3_3_2])

        next_layer = Concatenate()([x1_1,x3_2, x4_3, x2_4])

        return next_layer

    def block3_3(last_layer):
        x1_1 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=320, my_stried=1, my_padding='same')
        x1_2 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=384, my_stried=1, my_padding='same')
        x1_3 = my_conv2D(last_layer, my_kernel_size=(1, 1), my_filters=448, my_stried=1, my_padding='same')
        x1_4 = AveragePooling2D(pool_size=(3, 3), strides=1, padding='same')(last_layer)

        x2_2_1 = my_conv2D(x1_2, my_kernel_size=(1, 3), my_filters=384, my_stried=1, my_padding='same')
        x2_2_2 = my_conv2D(x1_2, my_kernel_size=(3, 1), my_filters=384, my_stried=1, my_padding='same')
        x2_3 = my_conv2D(x1_3, my_kernel_size=(3, 3), my_filters=384, my_stried=1, my_padding='same')
        x2_4 = my_conv2D(x1_4, my_kernel_size=(1, 1), my_filters=192, my_stried=1, my_padding='same')

        x3_2 = Concatenate()([x2_2_1, x2_2_2])
        x3_3_1 = my_conv2D(x2_3, my_kernel_size=(1, 3), my_filters=384, my_stried=1, my_padding='same')
        x3_3_2 = my_conv2D(x2_3, my_kernel_size=(3, 1), my_filters=384, my_stried=1, my_padding='same')

        x4_3 = Concatenate()([x3_3_1, x3_3_2])

        next_layer = Concatenate()([x1_1,x3_2, x4_3, x2_4])

        return next_layer


    input = Input(shape=input_shape,)
    layer1 = my_conv2D(input)
    layer2 = my_conv2D(layer1,my_stried=1)
    layer3 = my_conv2D(layer2,my_filters=64,my_stried=1)
    layer4 = MaxPooling2D((3,3),strides=2,padding='same')(layer3)
    layer5 = my_conv2D(layer4,my_filters=80,my_stried=1)
    layer6 = my_conv2D(layer5, my_filters=192, my_stried=1)
    layer7 = MaxPooling2D((3, 3), strides=2,padding='same')(layer6)

    layer1_1 = block1_1(layer7)
    layer1_2 = block1_2(layer1_1)
    layer1_3 = block1_3(layer1_2)

    layer2_1 = block2_1(layer1_3)
    layer2_2 = block2_2(layer2_1)
    layer2_3 = block2_3_4(layer2_2)
    layer2_4 = block2_3_4(layer2_3)
    layer2_5 = block2_5(layer2_4)

    layer3_1 = block3_1(layer2_5)
    layer3_2 = block3_2(layer3_1)
    layer3_3 = block3_3(layer3_2)


    inception_v3_model = tf.keras.Model(inputs=input,outputs=layer3_3)
    inception_v3_model.summary()
    return inception_v3_model

if __name__=='__main__':
	inception_v3_model = build_inception_v3()

三、待改进之处

• 有些代码重复度高，可增加函数参数将一类结构网络用一个函数实现
• 希望以后能熟练用类来搭建神经网络，这样显得专业<_<

四、学习心得

之前一直觉得自己搭建网络太难啦，想直接跑其他人的网络又看不懂代码，总觉得是自己不适合这个方向。之前看原理的时候感觉懂了，就是实现不来，便开始怀疑自己，后来发现还是框架不会用。于是花了点时间学习一些Keras函数详解，在慢慢跟着结构图上手感觉feel 倍儿棒！！加油，追梦人