import os
import cv2
import tqdmIMG_SIZE=224def ReadAndCutPicture(r_path,w_path):img = cv2.imread(r_path) #读到图片为BGRh=img.shape[0]w=img.shape[1]if(h<=w):new_h=IMG_SIZEnew_w=int(w/h*IMG_SIZE)img=cv2.resize(img,(new_w,new_h))  #输出图片尺寸为(宽，高)cut_w=int((new_w-new_h)/2)img=img[:,cut_w:cut_w+IMG_SIZE,:]else:new_w=IMG_SIZEnew_h=int(h/w*IMG_SIZE)img=cv2.resize(img,(new_w,new_h))cut_h=int((new_h-new_w)/2)img=img[cut_h:cut_h+IMG_SIZE,:,:]cv2.imwrite(w_path, img)normal_path=r'F:BaiduNetdiskDownloadmini-imagenetimages_normal' #裁剪后图片所在的位置
base_path=r'F:BaiduNetdiskDownloadmini-imagenetimages' #数据集所在位置for name in tqdm.tqdm(os.listdir(base_path)):r_path=os.path.join(base_path,name)w_dir=os.path.join(normal_path,name[:9])w_path=os.path.join(w_dir,name)if(os.path.exists(w_dir)==False):os.makedirs(w_dir)ReadAndCutPicture(r_path,w_path)

from keras.layers.core import Layer
from keras import backend as K
import tensorflow as tf
import numpy as np
class LRN(Layer):def __init__(self, alpha=0.0001,k=1,beta=0.75,n=5, **kwargs):self.alpha = alphaself.k = kself.beta = betaself.n = nsuper(LRN, self).__init__(**kwargs)def call(self, x, mask=None):b, ch, r, c = x.shapehalf_n = self.n // 2 # half the local region# orig keras code#input_sqr = T.sqr(x)  # square the inputinput_sqr = K.square(x) # square the input# orig keras code#extra_channels = T.alloc(0., b, ch + 2 * half_n, r,c)  # make an empty tensor with zero pads along channel dimension#input_sqr = T.set_subtensor(extra_channels[:, half_n:half_n+ch, :, :],input_sqr) # set the center to be the squared inputextra_channels = K.zeros((b, int(ch) + 2 * half_n, r, c))input_sqr = K.concatenate([extra_channels[:, :half_n, :, :],input_sqr, extra_channels[:, half_n + int(ch):, :, :]],axis = 1)scale = self.k # offset for the scalenorm_alpha = self.alpha / self.n # normalized alphafor i in range(self.n):scale += norm_alpha * input_sqr[:, i:i+int(ch), :, :]scale = scale ** self.betax = x / scalereturn xdef get_config(self):config = {"alpha": self.alpha,"k": self.k,"beta": self.beta,"n": self.n}base_config = super(LRN, self).get_config()return dict(list(base_config.items()) + list(config.items()))
class PoolHelper(Layer):def __init__(self, **kwargs):super(PoolHelper, self).__init__(**kwargs)def call(self, x, mask=None):return x[:,:,1:,1:]def get_config(self):config = {}base_config = super(PoolHelper, self).get_config()return dict(list(base_config.items()) + list(config.items()))

from keras.models import *
from keras.layers import *
from keras import models
from keras import initializers
from keras.utils import plot_model
normal_init=initializers.RandomNormal(mean=0.0, stddev=0.01, seed=None)
X=Input(shape=(224,224,3))
conv1=Conv2D(96,(11,11),strides=(4,4),padding='same',activation='relu',kernel_initializer=normal_init,bias_initializer='zeros',name='conv1')(X)
maxpooling1=MaxPooling2D(pool_size=(3,3),strides=2,name='maxpooling1')(conv1)
#lrn1=LRN(alpha=0.0001,k=2,beta=0.75,n=5)(maxpooling1)
conv2=Conv2D(256,(5,5),strides=(1,1),padding='same',activation='relu',kernel_initializer=normal_init,bias_initializer='zeros',name='conv2')(maxpooling1)
maxpooling2=MaxPooling2D(pool_size=(3,3),strides=2,name='maxpooling2')(conv2)
#lrn2=LRN(alpha=0.0001,k=2,beta=0.75,n=5)(maxpooling2)
conv3=Conv2D(384,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer=normal_init,bias_initializer='zeros',name='conv3')(maxpooling2)
conv4=Conv2D(384,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer=normal_init,bias_initializer='zeros',name='conv4')(conv3)
conv5=Conv2D(256,(3,3),strides=(1,1),padding='same',activation='relu',kernel_initializer=normal_init,bias_initializer='zeros',name='conv5')(conv4)
maxpooling3=MaxPooling2D(pool_size=(3,3),strides=2,name='maxpooling3')(conv5)
flatten=Flatten(name='flatten')(maxpooling3)
dense1=Dense(512,activation='relu',kernel_initializer=normal_init,name='dense1')(flatten)
dropout1=Dropout(0.5)(dense1)
dense2=Dense(512,activation='relu', kernel_initializer=normal_init,name='dense2')(dropout1)
dropout2=Dropout(0.5)(dense2)
outputs=Dense(100,activation='sigmoid',kernel_initializer=normal_init,name='output')(dropout2)
model = Model(X, outputs)
model.summary()
plot_model(model,to_file='AlexNet.png',show_shapes=True)

#定义数据生成器
from keras.preprocessing import image
from keras import optimizers
train_gen=image.ImageDataGenerator(featurewise_center=True,#输入数据数据减去数据集均值width_shift_range=0.2,#水平平移height_shift_range=0.2,#垂直平移horizontal_flip=True,#水平翻转brightness_range=[-0.1,0.1] #亮度变化范围
)
#编译模型
model.compile(loss='categorical_crossentropy',optimizer=optimizers.RMSprop(lr=1e-4),metrics=['acc'])
#训练图片路径
train_dir=r'F:BaiduNetdiskDownloadmini-imagenetimages_normal'
tg=train_gen.flow_from_directory(train_dir,target_size=(224,224),batch_size=128,class_mode='categorical'
)
tg.class_indices

history=model.fit_generator(tg,#训练数据生成器steps_per_epoch=10,#从生成器中抽取100个批次，epochs=5,#进行30轮训练shuffle=True,
)