1.输出结果
2.InceptionV4网络结构细节
https://mydreamambitious.blog.csdn.net/article/details/124414800
3.代码复现
import os
import keras
import numpy as np
import tensorflow as tf
from tensorflow.keras import models
from tensorflow.keras import layers
from tensorflow.keras.models import Model
#主干网络Stem
class Inception_ResNet_V2_Stem(tf.keras.Model):
def __init__(self):
super(Inception_ResNet_V2_Stem,self).__init__()
self.conv1=layers.Conv2D(32,kernel_size=[3,3],strides=[2,2],padding='valid')
self.conv1batch=layers.BatchNormalization()
self.conv2=layers.Conv2D(32,kernel_size=[3,3],strides=[1,1],padding='valid')
self.conv2batch=layers.BatchNormalization()
self.conv3=layers.Conv2D(64,kernel_size=[3,3],strides=[1,1],padding='same')
self.conv3batch=layers.BatchNormalization()
self.maxpool1=layers.MaxPool2D(pool_size=[3,3],strides=[2,2],padding='valid')
self.maxpool1conv4=layers.Conv2D(96,kernel_size=[3,3],strides=[2,2],padding='valid')
self.maxbatch1=layers.BatchNormalization()
self.conv5left=layers.Conv2D(64,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv5leftbatch=layers.BatchNormalization()
self.conv6left=layers.Conv2D(96,kernel_size=[3,3],strides=[1,1],padding='valid')
self.conv6leftbatch=layers.BatchNormalization()
self.conv5right=layers.Conv2D(64,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv5rightbatch=layers.BatchNormalization()
self.conv6right=layers.Conv2D(64,kernel_size=[7,1],strides=[1,1],padding='same')
self.conv6rightbatch=layers.BatchNormalization()
self.conv7right=layers.Conv2D(64,kernel_size=[1,7],strides=[1,1],padding='same')
self.conv7rightbatch=layers.BatchNormalization()
self.conv8right=layers.Conv2D(96,kernel_size=[3,3],strides=[1,1],padding='valid')
self.conv8rightbatch=layers.BatchNormalization()
self.maxpool2=layers.MaxPool2D(pool_size=[3,3],strides=[2,2],padding='valid')
self.conv9=layers.Conv2D(192,kernel_size=[3,3],strides=[2,2],padding='valid')
self.conv9batch=layers.BatchNormalization()
self.relu=layers.Activation('relu')
def call(self,inputs,training=None):
x_conv1=self.conv1(inputs)
x_conv1=self.conv1batch(x_conv1)
x_conv1=self.relu(x_conv1)
x_conv2=self.conv2(x_conv1)
x_conv2=self.conv2batch(x_conv2)
x_conv2=self.relu(x_conv2)
x_conv3=self.conv3(x_conv2)
x_conv3=self.conv3batch(x_conv3)
x_conv3=self.relu(x_conv3)
max_x=self.maxpool1(x_conv3)
max_conv_x=self.maxpool1conv4(x_conv3)
max_conv_x=self.maxbatch1(max_conv_x)
max_conv_x=self.relu(max_conv_x)
max_conv_out=tf.concat([
max_x,max_conv_x
],axis=3)
xleft=self.conv5left(max_conv_out)
xleft=self.conv5leftbatch(xleft)
xleft=self.relu(xleft)
xleft=self.conv6left(xleft)
xleft=self.conv6leftbatch(xleft)
xleft=self.relu(xleft)
xright=self.conv5right(max_conv_out)
xright=self.conv5rightbatch(xright)
xright=self.relu(xright)
xright=self.conv6right(xright)
xright=self.conv6rightbatch(xright)
xright=self.relu(xright)
xright=self.conv7right(xright)
xright=self.conv7rightbatch(xright)
xright=self.relu(xright)
xright=self.conv8right(xright)
xright=self.conv8rightbatch(xright)
xright=self.relu(xright)
x_left_right=tf.concat([
xleft,xright
],axis=3)
x_max_2=self.maxpool2(x_left_right)
x_conv9=self.conv9(x_left_right)
x_conv9=self.conv9batch(x_conv9)
x_conv9=self.relu(x_conv9)
x_max2_conv9=tf.concat([
x_max_2,x_conv9
],axis=3)
return x_max2_conv9
#InceptionV4_ResNet_V2的A模块
class InceptionV4_ResNet_V2_A(tf.keras.Model):
def __init__(self):
super(InceptionV4_ResNet_V2_A, self).__init__()
self.conv1=layers.Conv2D(32,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv1batch=layers.BatchNormalization()
self.conv2=layers.Conv2D(32,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv2batch=layers.BatchNormalization()
self.conv3=layers.Conv2D(32,kernel_size=[3,3],strides=[1,1],padding='same')
self.conv3batch=layers.BatchNormalization()
self.conv4=layers.Conv2D(32,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv4batch=layers.BatchNormalization()
self.conv5=layers.Conv2D(48,kernel_size=[3,3],strides=[1,1],padding='same')
self.conv5batch=layers.BatchNormalization()
self.conv6=layers.Conv2D(64,kernel_size=[3,3],strides=[1,1],padding='same')
self.conv6batch=layers.BatchNormalization()
self.conv1_3_6=layers.Conv2D(384,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv1_3_6batch=layers.BatchNormalization()
self.relu=layers.Activation('relu')
def call(self,inputs,training=None):
x1=self.conv1(inputs)
x1=self.conv1batch(x1)
x1=self.relu(x1)
x2=self.conv2(inputs)
x2=self.conv2batch(x2)
x2=self.relu(x2)
x2=self.conv3(x2)
x2=self.conv3batch(x2)
x2=self.relu(x2)
x3=self.conv4(inputs)
x3=self.conv4batch(x3)
x3=self.relu(x3)
x3=self.conv5(x3)
x3=self.conv5batch(x3)
x3=self.relu(x3)
x3=self.conv6(x3)
x3=self.conv6batch(x3)
x3=self.relu(x3)
x4=tf.concat([
x1,x2,x3
],axis=3)
x5=self.conv1_3_6(x4)
x5=self.conv1_3_6batch(x5)
x_out=tf.add(x5,inputs)
x_out = self.relu(x_out)
return x_out
#InceptionV4_ResNet_V2的B模块
class InceptionV4_ResNet_V2_B(tf.keras.Model):
def __init__(self):
super(InceptionV4_ResNet_V2_B, self).__init__()
self.conv1=layers.Conv2D(192,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv1batch=layers.BatchNormalization()
self.conv2=layers.Conv2D(128,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv2batch=layers.BatchNormalization()
self.conv3=layers.Conv2D(160,kernel_size=[1,7],strides=[1,1],padding='same')
self.conv3batch=layers.BatchNormalization()
self.conv4=layers.Conv2D(192,kernel_size=[7,1],strides=[1,1],padding='same')
self.conv4batch=layers.BatchNormalization()
self.relu=layers.Activation('relu')
self.conv5=layers.Conv2D(1152,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv5batch=layers.BatchNormalization()
def call(self,inputs,training=None):
print('np.shape: {}'.format(np.shape(inputs)))
x=self.conv1(inputs)
x=self.conv1batch(x)
x=self.relu(x)
x2=self.conv2(inputs)
x2=self.conv2batch(x2)
x2=self.relu(x2)
x2=self.conv3(x2)
x2=self.conv3batch(x2)
x2=self.relu(x2)
x2=self.conv4(x2)
x2=self.conv4batch(x2)
x2=self.relu(x2)
x3=tf.concat([
x,x2
],axis=3)
x4=self.conv5(x3)
x4=self.conv5batch(x4)
xx_out=tf.add(x4,inputs)
xx_out=self.relu(xx_out)
return xx_out
#InceptionV4_ResNet_V2的C模块
class InceptionV4_ResNet_V2_C(tf.keras.Model):
def __init__(self):
super(InceptionV4_ResNet_V2_C, self).__init__()
self.conv1=layers.Conv2D(192,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv1batch=layers.BatchNormalization()
self.conv2=layers.Conv2D(192,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv2batch=layers.BatchNormalization()
self.conv3=layers.Conv2D(224,kernel_size=[1,3],strides=[1,1],padding='same')
self.conv3batch=layers.BatchNormalization()
self.conv4=layers.Conv2D(256,kernel_size=[3,1],strides=[1,1],padding='same')
self.conv4batch=layers.BatchNormalization()
self.relu=layers.Activation('relu')
self.conv5=layers.Conv2D(2048,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv5batch=layers.BatchNormalization()
def call(self,inputs,training=None):
x=self.conv1(inputs)
x=self.conv1batch(x)
x=self.relu(x)
x2=self.conv2(inputs)
x2=self.conv2batch(x2)
x2=self.relu(x2)
x2=self.conv3(x2)
x2=self.conv3batch(x2)
x2=self.relu(x2)
x2=self.conv4(x2)
x2=self.conv4batch(x2)
x2=self.relu(x2)
x3=tf.concat([
x,x2
],axis=3)
x4=self.conv5(x3)
x4=self.conv5batch(x4)
xx_out=tf.add(x4,inputs)
xx_out=self.relu(xx_out)
return xx_out
#Inception-ResNet-V2的下采样A模块
class InceptionV4_ResNet_ReducetionModuleA(tf.keras.Model):
def __init__(self):
super(InceptionV4_ResNet_ReducetionModuleA, self).__init__()
self.maxpool=layers.MaxPool2D(pool_size=[3,3],strides=[2,2],padding='valid')
self.conv1=layers.Conv2D(384,kernel_size=[3,3],strides=[2,2],padding='valid')
self.conv1batch=layers.BatchNormalization()
self.conv2=layers.Conv2D(256,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv2batch=layers.BatchNormalization()
self.conv3=layers.Conv2D(256,kernel_size=[3,3],strides=[1,1],padding='same')
self.conv3batch=layers.BatchNormalization()
self.conv4=layers.Conv2D(384,kernel_size=[3,3],strides=[2,2],padding='valid')
self.conv4batch=layers.BatchNormalization()
self.relu=layers.Activation('relu')
def call(self,inputs,training=None):
maxpool=self.maxpool(inputs)
conv1=self.conv1(inputs)
conv1=self.conv1batch(conv1)
conv1=self.relu(conv1)
conv2=self.conv2(inputs)
conv2=self.conv2batch(conv2)
conv2=self.relu(conv2)
conv3=self.conv3(conv2)
conv3=self.conv3batch(conv3)
conv3=self.relu(conv3)
conv4=self.conv4(conv3)
conv4=self.conv4batch(conv4)
conv4=self.relu(conv4)
x_out=tf.concat([
conv1,maxpool,conv4
],axis=3)
return x_out
#Inception-ResNet-V1的下采样B模块
class InceptionV4_ResNet_ReducetionModuleB(tf.keras.Model):
def __init__(self):
super(InceptionV4_ResNet_ReducetionModuleB, self).__init__()
self.maxpool=layers.MaxPool2D(pool_size=[3,3],strides=[2,2],padding='valid')
self.conv1=layers.Conv2D(256,kernel_size=[1,1],strides=[1,1],padding='same')
self.conv1batch=layers.BatchNormalization()
self.conv2 = layers.Conv2D(384, kernel_size=[3,3], strides=[2,2], padding='valid')
self.conv2batch = layers.BatchNormalization()
self.conv3 = layers.Conv2D(256, kernel_size=[1, 1], strides=[1, 1], padding='same')
self.conv3batch = layers.BatchNormalization()
self.conv4 = layers.Conv2D(256, kernel_size=[3,3], strides=[2,2], padding='valid')
self.conv4batch = layers.BatchNormalization()
self.conv5 = layers.Conv2D(256, kernel_size=[1, 1], strides=[1, 1], padding='same')
self.conv5batch = layers.BatchNormalization()
self.conv6 = layers.Conv2D(256, kernel_size=[3,3], strides=[1, 1], padding='same')
self.conv6batch = layers.BatchNormalization()
self.conv7 = layers.Conv2D(256, kernel_size=[3, 3], strides=[2, 2], padding='valid')
self.conv7batch = layers.BatchNormalization()
self.relu=layers.Activation('relu')
def call(self,inputs,training=None):
max_x=self.maxpool(inputs)
x=self.conv1(inputs)
x=self.conv1batch(x)
x=self.relu(x)
x=self.conv2(x)
x=self.conv2batch(x)
x=self.relu(x)
x2=self.conv3(inputs)
x2=self.conv3batch(x2)
x2=self.relu(x2)
x2=self.conv4(x2)
x2=self.conv4batch(x2)
x2=self.relu(x2)
x3=self.conv5(inputs)
x3=self.conv5batch(x3)
x3=self.relu(x3)
x3=self.conv6(x3)
x3=self.conv6batch(x3)
x3=self.relu(x3)
x3=self.conv7(x3)
x3=self.conv7batch(x3)
x3=self.relu(x3)
x_out=tf.concat([
max_x,x,x2,x3
],axis=3)
return x_out
#整体网络结构
class Inception_ResNet_V2(tf.keras.Model):
def __init__(self,blocks,num_classes):
self.inception_resnet_V2_A = blocks[0]
self.inception_resnet_V2_B = blocks[1]
self.inception_resnet_V2_C = blocks[2]
super(Inception_ResNet_V2, self).__init__()
self.stem=self.Stem()
self.inceptionResNetA=self.inception_ResNet_V2_A(self.inception_resnet_V2_A)
self.inceptionReduceA=self.inception_Reduce_A()
self.inceptionResNetB = self.inception_ResNet_V2_B(self.inception_resnet_V2_B)
self.inceptionReduceB=self.inception_Reduce_B()
self.inceptionResNetC = self.inception_ResNet_V2_C(self.inception_resnet_V2_C)
self.avgpool=layers.GlobalAveragePooling2D()
self.dropout=layers.Dropout(rate=0.8)
self.dense=layers.Dense(num_classes)
self.softmax=layers.Activation('softmax')
def Stem(self):
stem=keras.Sequential([])
stem.add(Inception_ResNet_V2_Stem())
return stem
def inception_ResNet_V2_A(self,blocks):
inception_resnetA=keras.Sequential([])
for i in range(blocks):
inception_resnetA.add(InceptionV4_ResNet_V2_A())
return inception_resnetA
def inception_ResNet_V2_B(self,blocks):
inception_resnetB=keras.Sequential([])
for i in range(blocks):
inception_resnetB.add(InceptionV4_ResNet_V2_B())
return inception_resnetB
def inception_ResNet_V2_C(self,blocks):
inception_resnetC=keras.Sequential([])
for i in range(blocks):
inception_resnetC.add(InceptionV4_ResNet_V2_C())
return inception_resnetC
def inception_Reduce_A(self):
redA=keras.Sequential([])
redA.add(InceptionV4_ResNet_ReducetionModuleA())
return redA
def inception_Reduce_B(self):
redB=keras.Sequential([])
redB.add(InceptionV4_ResNet_ReducetionModuleB())
return redB
def call(self,inputs,training=None):
x=self.stem(inputs)
x=self.inceptionResNetA(x)
x=self.inceptionReduceA(x)
x = self.inceptionResNetB(x)
x = self.inceptionReduceB(x)
x = self.inceptionResNetC(x)
x=self.avgpool(x)
x=self.dropout(x)
x=self.dense(x)
x=self.softmax(x)
return x
model_inception_resnet_v2=Inception_ResNet_V2(blocks=[5,10,5],num_classes=1000)
model_inception_resnet_v2.build(input_shape=(None,299,299,3))
model_inception_resnet_v2.summary()
