03-08 06:58 阅读 120

keras实现Alexnet （cifar10数据集）

论文地址

网络结构

论文里画的有点乱，这个更清楚一点

cifar10数据集官网就可以直接下载，keras里也内置下载函数，用法和mnist类似

from tensorflow.keras.datasets import cifar10
from tensorflow.python.keras.utils import np_utils
import numpy as np
from tensorflow.keras import layers
from tensorflow.keras import models,optimizers
import matplotlib.pyplot as plt
import cv2

导入完头文件开始读取数据集，看一下数据集是什么样的

(x_train,y_train),(x_test,y_test)=cifar10.load_data()

print(x_train.shape)
print(y_train.shape)

y_train=np_utils.to_categorical(y_train)
y_test=np_utils.to_categorical(y_test)

#x_train=x_train.astype(‘float32‘)/255
#x_test=x_test.astype(‘float32‘)/255

print(y_train.shape,y_test.shape)

x_train_mean=np.mean(x_train)
x_test_mean=np.mean(x_test)

#x_train-=x_train_mean
#x_test-=x_test_mean

plt.imshow(x_train[0])
tt=np.zeros((224,224,3))
tt=cv2.resize(x_train[0],(224,224),interpolation=cv2.INTER_NEAREST)
plt.imshow(tt)

写网络模型

def Alexnet():
    inp=layers.Input(shape=(224,224,3))
    
    L1=layers.ZeroPadding2D((2,2))(inp)
    L1=layers.Conv2D(96,(11,11),strides=(4,4),padding=‘valid‘,activation=‘relu‘)(L1)#55
    L1=layers.MaxPooling2D((3,3),strides=(2,2))(L1)#27
    
    L2=layers.Conv2D(256,(5,5),activation=‘relu‘,padding=‘same‘)(L1)#27
    L2=layers.MaxPooling2D((3,3),strides=(2,2))(L2)#13
    
    L3=layers.Conv2D(384,(3,3),activation=‘relu‘,padding=‘same‘)(L2)#13
    
    L4=layers.Conv2D(384,(3,3),activation=‘relu‘,padding=‘same‘)(L3)#13
    
    L5=layers.Conv2D(256,(3,3),activation=‘relu‘,padding=‘same‘)(L4)#13
    L5=layers.MaxPooling2D((3,3),strides=(2,2))(L5)
    
    fc=layers.Flatten()(L5)
    
    fc1=layers.Dense(4096)(fc)
    fc1=layers.BatchNormalization()(fc1)
    fc1=layers.Activation(‘relu‘)(fc1)
    fc1=layers.Dropout(0.5)(fc1)
    
    fc2=layers.Dense(4096)(fc1)
    fc2=layers.BatchNormalization()(fc2)
    fc2=layers.Activation(‘relu‘)(fc2)
    fc2=layers.Dropout(0.5)(fc2)
    
    pred=layers.Dense(10)(fc2)
    pred=layers.BatchNormalization()(pred)
    pred=layers.Activation(‘softmax‘)(pred)
    
    model=models.Model(inp,pred)
    omz=optimizers.Adam(lr=0.01)
    model.compile(optimizer=omz,loss=‘categorical_crossentropy‘,metrics=[‘acc‘])
    model.summary()
    return model

cifar10数据集图像尺寸比较小，而alexnet的输入是224，所以要把样本都reshape，需要一点时间因为数据比较多

from tensorflow.keras.preprocessing import image
import cv2

X_TRAIN=np.zeros((x_train.shape[0],224,224,x_train.shape[3]))
X_TEST=np.zeros((x_test.shape[0],224,224,x_test.shape[3]))
print(X_TRAIN.shape,X_TEST.shape)
for i in range(x_train.shape[0]):
    X_TRAIN[i]=cv2.resize(x_train[i],(224,224),interpolation=cv2.INTER_NEAREST)
    
for i in range(x_test.shape[0]):
    X_TEST[i]=cv2.resize(x_test[i],(224,224),interpolation=cv2.INTER_NEAREST)

生成数据流

ImageDataGenerator使用方法见官方文档

train_data_gen=image.ImageDataGenerator()
train_gen=train_data_gen.flow(X_TRAIN,y_train,batch_size=32)

test_data_gen=image.ImageDataGenerator()
test_gen=test_data_gen.flow(X_TEST,y_test,batch_size=32)

开始训练

model1=Alexnet()
model1.fit_generator(train_gen,steps_per_epoch=50000/32,epochs=10)

50000是训练样本个数，也可以用x_train.shape[0]表示

模型结构：

训练结果

测试一下

model1.evaluate_generator(test_gen)

过拟合了，训练时加上验证集和callbacks.Earlystopping效果应该会改善

whatever，能用就行

顺便隔壁pytorch有官方版本，tf的没在github上找到

原文：https://www.cnblogs.com/wangtianning1223/p/15078364.html