GoogleNetV3详细结构图及其keras实现

目录

GoogleNetV3结构图

Keras完整代码

---

 

GoogleNetV3结构图

该结构图是由Google论文中给出的表格细化而来，程序是自己编写的，Injection层的输出shape与论文中给出的由细微差别，如有错漏请多多指教。

上图为论文中给出的结构表格，下图是keras.plot生成的详细结构：

 

Keras完整代码：

     将想要进行训练的图片放在对应文件夹中即可进行训练：

    例如    ./data/train/cat                             其中cat、dog为需要网络分辨的类别，存储在不同文件夹中flow_from_directory

               ./data/train/dog                            会自动对图片进行标记，有几个类别就建立几个文件夹。

               ./data/validation/cat                     train文件夹下的图片用于训练，validation文件夹下的图片用于验证准确性，关于

               ./data/validation/dog                  

               ImageDataGenerator的详细内容可参考：https://blog.csdn.net/caanyee/article/details/52502759 或

                                                                                 https://keras-cn.readthedocs.io/en/latest/

 

from keras.models import Model
from keras.layers import Input, Dense, Dropout, BatchNormalization, Flatten, concatenate, Activation, Reshape
from keras.layers.convolutional import Conv2D, MaxPooling2D, AveragePooling2D
from keras.preprocessing.image import ImageDataGenerator
from keras.utils import plot_model
from keras.utils import np_utils
import numpy as np
from time import ctime
from PIL import Image
import pickle
import glob
import os
import tensorflow as tf


def Conv2d_BN(x, nb_filter, kernel_size, padding='same', strides=(1, 1), name=None):
	if name is not None:
		bn_name = name + '_bn'
		conv_name = name + '_conv'
	else:
		bn_name = None
		conv_name = None

	x = Conv2D(nb_filter, kernel_size, padding=padding, strides=strides, name=conv_name)(x)
	x = BatchNormalization(axis=3, name=bn_name)(x)
	x = Activation('relu')(x)
	return x

def InceptionT1(x, nb_filter, strides = (1,1)):
	branch1 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)
	branch1 = Conv2d_BN(branch1, nb_filter, (3, 3), padding='same', strides=(1, 1), name=None)
	branch1 = Conv2d_BN(branch1, nb_filter, (3, 3), padding='same', strides=strides, name=None)


	branch2 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)
	branch2 = Conv2d_BN(branch2, nb_filter, (3, 3), padding='same', strides=strides, name=None)

	branch3 = MaxPooling2D(pool_size=(3, 3), strides=strides, padding='same')(x)
	branch3 = Conv2d_BN(branch3, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)


	branch4 = Conv2d_BN(x, nb_filter, (1, 1), padding='valid', strides=strides, name=None)
	# branch4 = tf.reshape(branch4,(17,17,192))

	x = concatenate([branch1, branch2, branch3, branch4], axis=3)

	return x

def InceptionT2(x, nb_filter, strides = (1,1), stride1 = (1, 1), stride2 = (1, 1)):
	branch1 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)
	branch1 = Conv2d_BN(branch1, nb_filter, (1, 7), padding='same', strides=(1, 1), name=None)
	branch1 = Conv2d_BN(branch1, nb_filter, (7, 1), padding='same', strides=(1, 1), name=None)
	branch1 = Conv2d_BN(branch1, nb_filter, (1, 7), padding='same', strides=stride1, name=None)
	branch1 = Conv2d_BN(branch1, nb_filter, (7, 1), padding='same', strides=stride2, name=None)


	branch2 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)
	branch2 = Conv2d_BN(branch2, nb_filter, (1, 7), padding='same', strides=stride1, name=None)
	branch2 = Conv2d_BN(branch2, nb_filter, (7, 1), padding='same', strides=stride2, name=None)

	branch3 = MaxPooling2D(pool_size=(3, 3), strides=strides, padding='same')(x)
	branch3 = Conv2d_BN(branch3, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)


	branch4 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=strides, name=None)

	x = concatenate([branch1, branch2, branch3, branch4], axis=3)

	return x

def InceptionT3(x, nb_filter, strides = (1,1)):
	branch1 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)
	branch1 = Conv2d_BN(branch1, nb_filter, (3, 3), padding='same', strides=strides, name=None)
	branch1_1 = Conv2d_BN(branch1, nb_filter, (1, 3), padding='same', strides=(1, 1), name=None)
	branch1_2 = Conv2d_BN(branch1, nb_filter, (3, 1), padding='same', strides=(1, 1), name=None)


	branch2 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=strides, name=None)
	branch2_1 = Conv2d_BN(branch2, nb_filter, (1, 3), padding='same', strides=(1, 1), name=None)
	branch2_2 = Conv2d_BN(branch2, nb_filter, (3, 1), padding='same', strides=(1, 1), name=None)

	branch3 = MaxPooling2D(pool_size=(3, 3), strides=strides, padding='same')(x)
	branch3 = Conv2d_BN(branch3, nb_filter, (1, 1), padding='same', strides=(1, 1), name=None)


	branch4 = Conv2d_BN(x, nb_filter, (1, 1), padding='same', strides=strides, name=None)

	x = concatenate([branch1_1,branch1_2, branch2_1, branch2_2, branch3, branch4], axis=3)

	return x

def creatcnn():
	inpt = Input(shape=(299, 299, 3))
	x = Conv2d_BN(inpt, 32, (3, 3), strides=(2, 2), padding='valid')#(149,149)32
	x = Conv2d_BN(x, 32, (3, 3), strides=(1, 1), padding='valid')#(147,147)32
	x = Conv2d_BN(x, 64, (3, 3), strides=(1, 1), padding='same')#(147,147)64
	x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), padding='valid')(x)#64
	x = Conv2d_BN(x, 80, (3, 3), strides=(1, 1), padding='valid')#80
	x = Conv2d_BN(x, 192, (3, 3), strides=(2, 2), padding='valid')#192
	x = Conv2d_BN(x, 288, (3, 3), strides=(1, 1), padding='same')#288
	x = InceptionT1(x, 192)  # 768
	x = InceptionT1(x, 192)  # 768
	x = InceptionT1(x, 192, strides = (2, 2))  # 768
	x = InceptionT2(x, 320)  # 1280
	x = InceptionT2(x, 320)  # 1280
	x = InceptionT2(x, 320)  # 1280
	x = InceptionT2(x, 320)  # 1280
	x = InceptionT2(x, 320, strides = (2, 2), stride1 = (2, 1), stride2 = (1, 2))  # 1280
	x = InceptionT3(x, 341)  # 2048
	x = InceptionT3(x, 341, strides = (1, 1))  # 2048
	x = AveragePooling2D(pool_size=(9, 9), strides=(9, 9), padding='same')(x)
	x = Flatten()(x)
	x = Dense(1000, activation='relu')(x)
	x = Dense(500, activation='relu')(x)
	x = Dense(1, activation='softmax')(x)
	model = Model(inpt, x, name='inception')
	return model

if __name__ == "__main__":

	model = creatcnn()
	plot_model(model, to_file='GoogleNetV3_model.png', show_shapes=True)
	model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

	train_datagen = ImageDataGenerator(
		shear_range=0.2,
		zoom_range=0.2,
		horizontal_flip=True)

	test_datagen = ImageDataGenerator()

	train_generator = train_datagen.flow_from_directory(
		'data/train',
		target_size=(299, 299),
		batch_size=10,
		class_mode='binary')

	validation_generator = test_datagen.flow_from_directory(
		'data/validation',
		target_size=(299, 299),
		batch_size=10,
		class_mode='binary')

	model.fit_generator(
		train_generator,
		steps_per_epoch=97,
		epochs=50,
		validation_data=validation_generator,
		validation_steps=7,
		verbose=2)
	model.save_weights('GoogleNet_weight.h5')