[经典阅读]Alexnet 论文读书笔记以及代码复现

Alexnet主要内容

1.使用数据集为ImageNet
2.使用非饱和激活函数Relu比饱和激活函数具有更快的收敛速度
3.使用两个GPU并行训练
4.使用Local Response Normalization进行优化
5.网络结构：5层卷积层，3层全连接层，1，2，5层卷积层后接最大池化
，使用的是重叠池化令步长s<窗口边长z,这里令s=2,z=3，每层都使用Relu激活
6.减少过拟合使用两种方法，1、数据增广 2、Dropout
7.实现细节：使用随机梯度下降，动量参数为0.9，权值和偏置的初始化策略

复现Alexnet

做出的一些改动:
1.数据集的变化：ImageNet–>cifar10，图像大小由2242243–>32323
2.没有使用原文提到的双GPU策略，没有使用局部影响归一化策略
使用平台：
Google Colab 提供NVIDIA Tesla K80 GPU免费使用，不过存在一些限制，连接时间不超过12小时，Colab Pro版本收费且只提供给美国地区使用

代码部分

import keras
from keras.datasets import cifar10
from keras import backend as K
from keras.layers import Input, Conv2D, GlobalAveragePooling2D, Dense, BatchNormalization, Activation, MaxPooling2D
from keras.models import Model
from keras.layers import concatenate,Dropout,Flatten

from keras import optimizers,regularizers
from keras.preprocessing.image import ImageDataGenerator
from keras.initializers import he_normal
from keras.callbacks import LearningRateScheduler, TensorBoard, ModelCheckpoint

num_classes = 10
batch_size = 64
iterations =782
epochs = 300
DROPOUT = 0.5
CONCAT_AXIS = 3
weight_decay = 1e-4
DATA_FORMAT = 'channels_last'
log_filepath = './alexnet'


def color_preprocessing(x_train,x_test):
	x_train = x_train.astype('float32')
	x_test = x_test.astype('float32')
	mean = [125.307,122.95,113.865]
	std = [62.9932,62.0887,66.7048]
	for i in range(3):
		x_train[:,:,:,i] = (x_train[:,:,:,i]-mean[i]) / std[i]
		x_test[:,:,:,i] = (x_test[:,:,:,i]-mean[i]) / std[i]
	return x_train,x_test
def scheduler(epoch):
	if epoch < 100:
		return 0.01
	if epoch < 200:
		return 0.001
	return 0.0001

#load_data
(x_train, y_train),(x_test, y_test) = cifar10.load_data()
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test  = keras.utils.to_categorical(y_test, num_classes)
x_train,x_test = color_preprocessing(x_train, x_test)


def alexnet(img_input,classes=10):
	#1
	x = Conv2D(96, (11,11), strides = (4,4), padding = 'same',
		activation = 'relu', kernel_initializer = 'uniform')(img_input)#may kernel_initializer need change
	#11*11*3*96+96= 34944
	x = MaxPooling2D(pool_size = (3,3), strides = (2,2), padding = 'same', data_format = DATA_FORMAT)(x)
	
	#2
	x = Conv2D(256,(5,5), strides = (1,1), padding = 'same',
		activation = 'relu', kernel_initializer = 'uniform')(x)
	#5*5*96*256+256 = 614656
	x = MaxPooling2D(pool_size = (3,3), strides = (2,2), padding = 'same', data_format = DATA_FORMAT)(x)

	#3
	x = Conv2D(384, (3,3), strides = (1,1), padding = 'same',
		activation = 'relu', kernel_initializer = 'uniform')(x)
	#3*3*256*384+384=885120
	#4
	x = Conv2D(384, (3,3), strides = (1,1), padding = 'same',
		activation = 'relu', kernel_initializer = 'uniform')(x)
	#3*3*384*384+384=1327488
	#5
	x = Conv2D(256, (3,3), strides = (1,1), padding = 'same',
		activation = 'relu', kernel_initializer = 'uniform')(x)
	#3*3*384*256+256=884992

	x = MaxPooling2D(pool_size = (3,3), strides = (2,2), padding = 'same', data_format = DATA_FORMAT)(x)		

	x = Flatten()(x)
	x = Dense(4096,activation = 'relu')(x)
	#256*4096+4096=1052672 
	x = Dropout(0.5)(x)
	x = Dense(4096,activation = 'relu')(x)
	#4096*4096+4096 = 16781312
	x = Dropout(0.5)(x)
	out = Dense(classes, activation = 'softmax')(x)
	#4096*10+10 = 40970

	return out

img_input = Input(shape = (32,32,3))
output = alexnet(img_input)
model = Model(img_input, output)
model.summary()

sgd = optimizers.SGD(lr = .1,momentum = 0.9 ,nesterov = True)
model.compile(loss = 'categorical_crossentropy',optimizer = sgd, metrics = ['accuracy'])
tb_cb = TensorBoard(log_dir = log_filepath,histogram_freq = 0)
change_lr = LearningRateScheduler(scheduler)
cbks = [change_lr, tb_cb]

datagen = ImageDataGenerator(horizontal_flip  = True,
							 width_shift_range = 0.125,
							 height_shift_range = 0.125,
							 fill_mode = 'constant',cval = 0.)
datagen.fit(x_train)

#train 
model.fit_generator(datagen.flow(x_train,y_train,batch_size = batch_size),
	steps_per_epoch = iterations,
	epochs = epochs,
	callbacks = cbks,
	validation_data = (x_test,y_test))

model.save('alexnet.h5')
print('sucess!')

实验结果