本文共 4422 字，大约阅读时间需要 14 分钟。

神经网络比较深…下面的代码最好运行在GPU上

环境参数：Keras == 2.1.2

import kerasfrom keras.datasets import cifar10from keras.preprocessing.image import ImageDataGeneratorfrom keras.models import Sequentialfrom keras.layers import Dense,Dropout,Flatten,Activationfrom keras.layers import Conv2D,MaxPooling2D,ZeroPadding2D,GlobalMaxPooling2D#加载数据集batch_size = 32num_classes = 10epochs = 1600data_augmentation = True(x_train,y_train),(x_test,y_test) = cifar10.load_data()print('x_train shape:',x_train.shape)print(x_train.shape[0],'train samples')print(x_test.shape[0],'test samples')x_train = x_train.astype('float32')x_test  = x_test.astype('float32')x_train /= 255x_test  /= 255y_train =keras.utils.to_categorical(y_train,num_classes) y_test  =keras.utils.to_categorical(y_test,num_classes)#搭建网络 model = Sequential()model.add(Conv2D(32, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(32, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(32, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(48, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(48, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(MaxPooling2D(pool_size=(2, 2)))model.add(Dropout(0.25))model.add(Conv2D(80, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(80, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(80, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(80, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(80, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(MaxPooling2D(pool_size=(2, 2)))model.add(Dropout(0.25))model.add(Conv2D(128, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(128, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(128, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(128, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(Conv2D(128, (3, 3), padding='same',input_shape=x_train.shape[1:]))model.add(Activation('relu'))model.add(GlobalMaxPooling2D())model.add(Dropout(0.25))model.add(Dense(500))model.add(Activation('relu'))model.add(Dropout(0.25))model.add(Dense(num_classes))model.add(Activation('softmax'))model.summary()#模型编译训练opt = keras.optimizers.Adam(lr = 0.0001)model.compile(loss='categorical_crossentropy',optimizer = opt,metrics = ['accuracy'])print("---------train---------")model.fit(x_train,y_train,epochs = 600,batch_size = 128,)print("---------test---------")loss,acc = model.evaluate(x_test,y_test)print("loss=",loss)print("accuracy=",acc)#基于数据增强的训练方法if not data_augmentation:   print('Not using data augmentation.')   model.fit(x_train, y_train,             batch_size=batch_size,             epochs=epochs,             validation_data=(x_test, y_test),             shuffle=True, callbacks=[tbCallBack])else:   print('Using real-time data augmentation.')   datagen = ImageDataGenerator(       featurewise_center=False,  # set input mean to 0 over the dataset       samplewise_center=False,  # set each sample mean to 0       featurewise_std_normalization=False,  # divide inputs by std of the dataset       samplewise_std_normalization=False,  # divide each input by its std       zca_whitening=False,  # apply ZCA whitening       rotation_range=10,  # randomly rotate images in the range (degrees, 0 to 180)       width_shift_range=0.2,  # randomly shift images horizontally (fraction of total width)       height_shift_range=0.2,  # randomly shift images vertically (fraction of total height)       horizontal_flip=True,  # randomly flip images       vertical_flip=False)  # randomly flip images   datagen.fit(x_train)   model.fit_generator(datagen.flow(x_train,y_train,batch_size=batch_size),                                                          steps_per_epoch=x_train.shape[0] // batch_size,                       epochs=epochs,                       validation_data=(x_test, y_test), callbacks=[tbCallBack])

转载地址：http://ydkab.baihongyu.com/