交叉验证是一种用于评估深度学习模型性能的统计方法。交叉验证是一种重采样方法，用于在有限的数据样本上评估深度学习模型，可用于分类、回归等任务。交叉验证可以减少过拟合、提供稳健的性能评估、高效利用数据(尤其是在数据量有限的情况下)。
K-Fold Cross-Validation：K折交叉验证，K表示给定数据样本随机要分成的大小相等的组数(折)。训练K次，将模型在所有K次迭代中的性能取平均值，从而得到模型泛化能力的估计。如下图所示：原图来自于： https://www.kaggle.com
一般步骤：数据集中的每个样本有且仅一次用于测试，每个样本均K-1次用于训练
1. 随机打乱数据集，将数据集分成K折。
2. 训练及测试：对于每个折，使用K-1个折来训练模型，使用剩余的折作为测试集来评估模型。
3. 汇总结果：计算每个折的性能指标，并取平均值。
K值的选择会影响偏差和方差之间的权衡，K的选择通常为5或10，没有硬性规定。小数据集适合较大的K值，中数据集适合较小的K值。大数据集不建议使用K-Fold。
1. 较小的K值，使计算速度更快，但性能估计的方差更大。
2. 较大的K值，使方差更小，但计算成本更高。
以下测试代码为将数据集按照K折交叉验证拆分，并计算mean和std，用于回归训练中，csv文件中为每幅图像对应一个float值
def parse_args(): parser = argparse.ArgumentParser(description="K-Fold Cross-Validation") parser.add_argument("--src_dataset_path", required=True, type=str, help="source dataset path") parser.add_argument("--src_csv_file", required=True, type=str, help="source csv file") parser.add_argument("--dst_dataset_path", required=True, type=str, help="the path of the destination dataset after split") parser.add_argument("--k", type=int, default=5, help="number fo groups, K-Fold cross validataion") args = parser.parse_args() return args def split_k_fold(src_dataset_path, dst_dataset_path, src_csv_file, k): if src_dataset_path is None or not src_dataset_path or not Path(src_dataset_path).is_dir(): raise ValueError(colorama.Fore.RED + f"{src_dataset_path} is not a directory") if src_csv_file is None or not src_csv_file or not Path(src_csv_file).is_file(): raise ValueError(colorama.Fore.RED + f"{src_csv_file} is not a file") for i in range(1, k+1): path_name = dst_dataset_path + "_" + str(i) if Path(path_name).exists(): raise FileExistsError(colorama.Fore.RED + f"specified directory already exists: {path_name}") Path(path_name).mkdir(parents=True) Path(path_name + "/train").mkdir(parents=True) Path(path_name + "/val").mkdir(parents=True) dataframe = pd.read_csv(src_csv_file, header=None) samples = dataframe.values.tolist() if len(samples) == 0: raise FileNotFoundError(colorama.Fore.RED + f"there is no data in the file: {src_csv_file}") print(f"samples length: {len(samples)}; samples0: {samples[0]}") images = [img for img in Path(src_dataset_path).glob("*") if img.is_file()] if len(images) == 0: raise FileNotFoundError(colorama.Fore.RED + f"there are no matching images in this directory: {src_dataset_path}") print(f"images number: {len(images)}, image0: {images[0]}") if len(samples) != len(images): raise ValueError(colorama.Fore.RED + f"length mismatch: samples:{len(samples)}; images:{len(images)}") for i in range(0, len(samples)): if samples[i][2] != images[i].name: raise ValueError(colorama.Fore.RED + f"name mismatch: samples{i}:{samples[i][2]}; images{i}:{images[i].name}") total = len(samples) numbers = list(range(total)) random.shuffle(numbers) fold_size = total // k def write_and_copy(index_list, folder, csv_file): with open(csv_file, mode="w", newline="", encoding="utf-8") as file: write = csv.writer(file) for idx in index_list: write.writerow(samples[idx]) shutil.copy(images[idx], folder) def calculate_mean_std(train_list, txt_file): mean_sum = np.zeros(3, dtype=np.float64) std_sum = np.zeros(3, dtype=np.float64) for idx in train_list: img = cv2.imread(str(images[idx])) if img is None: raise FileNotFoundError(colorama.Fore.RED + f"image file does not exist: {images[idx]}") img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) img = img.astype(np.float64) / 255.0 mean_sum += img.mean(axis=(0, 1)) std_sum += img.std(axis=(0, 1)) mean = mean_sum / len(train_list) std = std_sum / len(train_list) with open(txt_file, mode="w", encoding="utf-8") as file: file.write(f'--mean "{tuple(mean.tolist())}" --std "{tuple(std.tolist())}"') for fold in range(k): val_idx = numbers[fold * fold_size : (fold + 1) * fold_size] if fold < k - 1 else numbers[fold * fold_size :] train_idx = [i for i in numbers if i not in val_idx] print(f"fold {fold + 1}: length val: {len(val_idx)}; train: {len(train_idx)}") write_and_copy(train_idx, dst_dataset_path + "_" + str(fold+1) + "/train", dst_dataset_path + "_" + str(fold+1) + "/train.csv") write_and_copy(val_idx, dst_dataset_path + "_" + str(fold+1) + "/val", dst_dataset_path + "_" + str(fold+1) + "/val.csv") calculate_mean_std(train_idx, dst_dataset_path + "_" + str(fold+1) + "/mean_std.txt") if __name__ == "__main__": colorama.init(autoreset=True) args = parse_args() split_k_fold(args.src_dataset_path, args.dst_dataset_path, args.src_csv_file, args.k) print(colorama.Fore.GREEN + "====== execution completed ======")执行结果如下图所示：
GitHub：https://github.com/fengbingchun/NN_Test
