keras學習記錄——訓練resnet模型對cifar10分類

目錄

keras學習記錄——訓練resnet模型對cifar10分類

前言

一、資料準備

二、構建網路

三、網路訓練

四、訓練結果

總結

前言

最近在對resnet中的add操作做量化處理，所以先將resnet對cifar10分類訓練記錄下來，

提示：以下是本篇文章正文內容，下面案例可供參考

一、資料準備

1. 直接呼叫keras中的資料庫：

   from keras.datasets import cifar10
   (data_train, label_train), (data_test, label_test) = cifar10.load_data()
   data_train = (data_train-127.5) / 128
   data_test = (data_test - 127.5) /128

   如果需要將便簽轉變為one-hot編碼，可以使用如下：

   label_train = keras.utils.to_categorical(label_train,10)
   label_test = keras.utils.to_categorical(label_test, 10)

2. 如果呼叫時太慢，可以下載cifar-10-python.tar.gz，網上有很多資源，這里就不放鏈接了，之后將cifar10.load_data()函式中下述路徑改為本地路徑即可，

   origin = 'https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz'

二、構建網路

1. resnet_block

代碼如下：

def resnet_block(input, num_filters=16, kernel_size=3, strides=1, activation='relu'):
    # x = Conv2D(num_filters, kernel_size=kernel_size,strides=strides,padding='same')(input)
    x = Conv2D(num_filters, kernel_size=kernel_size, strides=strides, padding='same',
           kernel_initializer='he_normal', kernel_regularizer=l2(1e-4))(input)
    x = BatchNormalization()(x)
    x = Activation('relu')(x)
    x = Conv2D(num_filters,  kernel_size=kernel_size, strides=1,padding='same')(x)
    x = BatchNormalization()(x)
    if strides==1:
        shortcut = input
    else:
        shortcut = Conv2D(num_filters,  kernel_size=kernel_size,strides=strides,padding='same')(input)
    # x = myAdd(activation='relu')([x, shortcut])
    x = keras.layers.add([x, shortcut])
    x = Activation('relu')(x)
    return x

結構如下：

2. resnet

代碼如下：

def resnet(input_shape):
    inputs = Input(shape=input_shape)

    # 1
    x = Conv2D(16, activation='relu',  kernel_size=3, strides=1,padding='same')(inputs)
    x = BatchNormalization()(x)
    print('layer1,xshape:', x.shape)
    # 2-7
    for i in range(6):
        x = resnet(x, num_filters=16, kernel_size=3, strides=1, activation='relu')
    # out: 32*32*16
    # 8-13
    for i in range(6):
        if i == 0:
            x = resnet(x, num_filters=32, kernel_size=3, strides=2, activation='relu')
        else:
            x = resnet(x, num_filters=32, kernel_size=3, strides=1, activation='relu')

    # out: 16*16*32
    # 14-19
    for i in range(6):
        if i == 0:
            x = resnet(x, num_filters=64, kernel_size=3, strides=2, activation='relu')
        else:
            x = resnet(x, num_filters=64, kernel_size=3, strides=1, activation='relu')

    # out: 8*8*64
    x = AveragePooling2D(pool_size=2)(x)
    # out: 4*4*64
    y = Flatten()(x)
    # out:1024
    outputs = Dense(10, kernel_initializer='he_normal')(y)
    # outputs = myDense(10,  kernel_initializer='glorot_normal', use_batchnormal = False)(y)

    outputs = Activation('softmax')(outputs)
    model = Model(inputs=inputs, outputs=outputs)
    return model

三、網路訓練

代碼如下：

model = resnet((32,32,3))
model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(), metrics=['accuracy'])
model.summary()
plot_model(model, to_file='./model.png', show_shapes=True)
# checkpoint = ModelCheckpoint(filepath='saved_models/'+config['mode']+'-'+str(config['epoch'])+'-'+str(config['batch_size'])+'-{val_accuracy:.4f}'+'.h5',monitor='val_accuracy',
#                              verbose=1,save_best_only=True)
csv_logger = CSVLogger('saved_models/'+config['mode']+'-'+str(config['epoch'])+'-'+str(config['batch_size'])+'.log', append=False)
lr_reducer = ReduceLROnPlateau(monitor='val_accuracy', factor=0.2, patience=5,
                               mode='max', min_lr=1e-3)

# model.fit(data_train, label_train,batch_size=64,epochs=1, validation_data=(data_test,label_test), verbose=1, callbacks=callbacks)
aug = ImageDataGenerator(width_shift_range=0.2,height_shift_range=0.2,horizontal_flip=True,zoom_range=0.2)
aug.fit(data_train)
gen = aug.flow(data_train, label_train, batch_size=config['batch_size'])
model.fit_generator(generator=gen,epochs=config['epoch'],validation_data=(data_test,label_test),
                    callbacks=[lr_reducer, csv_logger], verbose=1)

四、訓練結果

由于時間關系，并未將網路訓練完全，

總結

學習到了關于keras中訓練的小技巧，如：earlystop等，并且學會了如何構建殘差神經網路，