classification-models-tensorflow/train_vgg16.py

import os

import tensorflow as tf
from keras.callbacks import ModelCheckpoint, CSVLogger
from tensorflow.keras.preprocessing.image import ImageDataGenerator

from models.VGG16 import create_model


def load_data(train_dir, val_dir, img_size=(224, 224), batch_size=32):
    # 给定的均值和标准差
    mean = tf.constant([0.485, 0.456, 0.406])
    std = tf.constant([0.229, 0.224, 0.225])

    # 自定义标准化函数
    def normalize_image(x):
        return (x - mean) / std  # 标准化

    # 使用 ImageDataGenerator 加载图像并进行预处理
    train_datagen = ImageDataGenerator(
        rescale=1.0 / 255.0,  # 归一化
        rotation_range=40,
        width_shift_range=0.2,
        height_shift_range=0.2,
        shear_range=0.2,
        zoom_range=0.2,
        horizontal_flip=True,
        fill_mode='nearest',
        preprocessing_function=normalize_image  # 使用自定义的标准化函数
    )

    val_datagen = ImageDataGenerator(
        rescale=1.0 / 255.0,
        preprocessing_function=normalize_image  # 使用自定义的标准化函数
    )

    train_generator = train_datagen.flow_from_directory(
        train_dir,
        target_size=img_size,  # VGG16 输入大小
        batch_size=batch_size,
        class_mode='categorical'  # 多分类任务
    )

    val_generator = val_datagen.flow_from_directory(
        val_dir,
        target_size=img_size,
        batch_size=batch_size,
        class_mode='categorical'
    )

    return train_generator, val_generator


def find_latest_checkpoint(directory):
    # 获取指定目录下的所有 .h5 文件
    checkpoint_files = [f for f in os.listdir(directory) if f.endswith('.h5')]
    if not checkpoint_files:
        return None
    # 按照文件名中的数字进行排序，找到最新的 epoch 文件
    checkpoint_files.sort(key=lambda x: int(x.split('_')[-1].split('.')[0]))
    return os.path.join(directory, checkpoint_files[-1])


def train_model(args, train_generator, val_generator):
    # Create model
    model = create_model()

    # 调整学习率
    learning_rate = args.lr if args.lr else 1e-2

    # Select optimizer based on args.opt
    if args.opt == 'sgd':
        optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate,
                                            momentum=args.momentum if args.momentum else 0.0)
    elif args.opt == 'adam':
        optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
    else:
        optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)  # Default to Adam if unspecified

    # 编译模型
    model.compile(optimizer=optimizer,
                  loss='categorical_crossentropy',
                  metrics=['accuracy'])

    # Check if a checkpoint exists and determine the initial_epoch
    latest_checkpoint = find_latest_checkpoint(args.output_dir)
    if latest_checkpoint:
        model.load_weights(latest_checkpoint)  # Load the weights from the checkpoint
        initial_epoch = int(latest_checkpoint.split('_')[-1].split('.')[0])  # Get the last epoch from filename
        print(f"Resuming training from epoch {initial_epoch}")
    else:
        initial_epoch = 0
        print("No checkpoint found. Starting training from scratch.")

    # Define CSVLogger to log training history to a CSV file
    csv_logger = CSVLogger(os.path.join(args.output_dir, 'training_log.csv'), append=True)

    # Define ModelCheckpoint callback to save weights for each epoch
    checkpoint_callback = ModelCheckpoint(
        os.path.join(args.output_dir, 'vgg16_{epoch:03d}.h5'),  # Save weights as vgg16_{epoch}.h5
        save_weights_only=False,
        monitor='val_loss',  # Monitor the validation loss
        save_freq='epoch',  # Save after every epoch
        verbose=1
    )

    # 训练模型
    history = model.fit(
        train_generator,
        steps_per_epoch=train_generator.samples // train_generator.batch_size,
        epochs=args.epochs,
        validation_data=val_generator,
        validation_steps=val_generator.samples // val_generator.batch_size,
        initial_epoch=initial_epoch,
        callbacks=[csv_logger, checkpoint_callback]
    )

    return history


def get_args_parser(add_help=True):
    import argparse

    parser = argparse.ArgumentParser(description="PyTorch Classification Training", add_help=add_help)

    parser.add_argument("--data-path", default="dataset/imagenette2-320", type=str, help="dataset path")
    parser.add_argument("--output-dir", default="checkpoints/vgg16", type=str, help="path to save outputs")

    parser.add_argument(
        "-b", "--batch-size", default=2, type=int, help="images per gpu, the total batch size is $NGPU x batch_size"
    )
    parser.add_argument("--epochs", default=90, type=int, metavar="N", help="number of total epochs to run")

    parser.add_argument("--opt", default="sgd", type=str, help="optimizer")
    parser.add_argument("--lr", default=0.01, type=float, help="initial learning rate")
    parser.add_argument("--momentum", default=0.9, type=float, metavar="M", help="momentum")

    parser.add_argument(
        "--input-size", default=224, type=int, help="the random crop size used for training (default: 224)"
    )
    return parser


if __name__ == "__main__":
    args = get_args_parser().parse_args()

    # Set directories for your custom dataset
    train_dir = os.path.join(args.data_path, "train")
    val_dir = os.path.join(args.data_path, "val")

    # Set the directory where you want to save weights
    os.makedirs(args.output_dir, exist_ok=True)

    # Load data
    train_generator, val_generator = load_data(train_dir, val_dir, img_size=(args.input_size, args.input_size),
                                               batch_size=args.batch_size)

    # Start training
    train_model(args, train_generator, val_generator)