十一. 迁移学习

import tensorflow as tf
import glob as glob

def load_process_image(path,label):
    image = tf.io.read_file(path)
    image = tf.image.decode_jpeg(image,channels=3)
    image = tf.image.resize(image,[256,256])
    image = image/255
    return image,label

if __name__ == '__main__':
    train_image_path = glob.glob('.\dc_2000\\train\*\*.jpg')#['.', 'dc_2000', 'train', 'cat', 'cat.0.jpg']
    train_image_label = [int(p.split('\\')[3] == 'cat') for p in train_image_path]#1为cat
    train_image_ds = tf.data.Dataset.from_tensor_slices((train_image_path,train_image_label))

    AUTOTUNE = tf.data.experimental.AUTOTUNE#map并行处理
    train_image_ds = train_image_ds.map(load_process_image, num_parallel_calls=AUTOTUNE)
    BATCH_SIZE = 32
    train_count = len(train_image_path)
    train_image_ds = train_image_ds.shuffle(train_count).repeat().batch(BATCH_SIZE)

    test_image_path = glob.glob('.\dc_2000\\train\*\*.jpg')
    test_image_label = [int(p.split('\\')[3] == 'cat') for p in train_image_path]
    test_image_ds = tf.data.Dataset.from_tensor_slices((test_image_path, test_image_label))
    test_image_ds = test_image_ds.map(load_process_image, num_parallel_calls=AUTOTUNE)
    test_image_ds = test_image_ds.repeat().batch(BATCH_SIZE)
    test_count = len(test_image_path)

    #keras内置经典网络
    covn_base = tf.keras.applications.xception.Xception(
        weights='imagenet',#使用在imagenet训练好权重
        include_top=False,#只引入卷积基，不引入输出层
        input_shape=(256,256,3),
        pooling='avg'
    )
    model = tf.keras.Sequential()
    model.add(covn_base)
    model.add(tf.keras.layers.Dense(512,activation='relu'))
    model.add(tf.keras.layers.Dense(1, activation='sigmoid'))

    #covn_base.trainable = False,xception需要可训练

    model.compile(
        optimizer='adam',
        loss=tf.keras.losses.binary_crossentropy,
        metrics=['acc']
    )

    history = model.fit(
        train_image_ds,
        steps_per_epoch=train_count//BATCH_SIZE,
        epochs=5,
        validation_data=test_image_ds,
        validation_steps=test_count//BATCH_SIZE
    )
    print(history)

微调
只有分类器已经训练好了，才能微调卷积基的顶部卷积层。如果有没有这样的话，刚开始的训练误差很大，微调之前这些卷积层学到的表示会被破坏掉

1. 在预训练卷积基上添加自定义层
2. 冻结卷积基所有层
3. 训练添加的分类层
4. 解冻卷积基的一部分层
5. 联合训练解冻的卷积层和添加的自定义层

covn_base.trainable = True
    for layer in covn_base.layers[:-33]:
        layer.trainable=False#对后33个层训练