PyTorch多数据集合并问题：自定义韩文数据集路径错误致文件找不到

问题 #

训练一款可识别手写韩文、英文及数字的AI模型时，需要合并自定义韩文数据集、MJSynth和SynthText三个数据集。但合并后，train_set会默认使用MJSynth的路径，导致自定义数据集中的图片（如긴장_1227682.jpg）被错误地到MJSynth目录下查找，引发FileNotFoundError。

代码

custom_train_set = RecognitionDataset(
            parts[0].joinpath("images"),
            parts[0].joinpath("labels.json"),
            img_transforms=Compose(
                [
                    T.Resize((args.input_size, 4 * args.input_size), preserve_aspect_ratio=True),
                    # Augmentations
                    T.RandomApply(T.ColorInversion(), 0.1),
                    ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.02),
                ]
            ),
        )
        if len(parts) > 1:
            for subfolder in parts[1:]:
                custom_train_set.merge_dataset(
                    RecognitionDataset(subfolder.joinpath("images"), subfolder.joinpath("labels.json"))
                )

        train_set = MJSynth(
            train=True,
            img_folder='/media/cvpr/CM_22/mjsynth/mnt/ramdisk/max/90kDICT32px',
            label_path='/media/cvpr/CM_22/mjsynth/mnt/ramdisk/max/90kDICT32px/imlist.txt',
            img_transforms=T.Resize((args.input_size, 4 * args.input_size), preserve_aspect_ratio=True),
        )

        _train_set = SynthText(
            train=True,
            recognition_task=True,
            download=True,  # NOTE: download can take really long depending on your bandwidth
            img_transforms=T.Resize((args.input_size, 4 * args.input_size), preserve_aspect_ratio=True),
        )
        train_set.data.extend([(np_img, target) for np_img, target in _train_set.data])
        train_set.data.extend([(np_img, target) for np_img, target in custom_train_set.data])

报错信息

Traceback (most recent call last):
  File "/media/cvpr/CM_22/doctr/references/recognition/train_pytorch.py", line 485, in <module>
    main(args)
  File "/media/cvpr/CM_22/doctr/references/recognition/train_pytorch.py", line 396, in main
    fit_one_epoch(model, train_loader, batch_transforms, optimizer, scheduler, mb, amp=args.amp)
  File "/media/cvpr/CM_22/doctr/references/recognition/train_pytorch.py", line 118, in fit_one_epoch
    for images, targets in progress_bar(train_loader, parent=mb):
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/fastprogress/fastprogress.py", line 50, in __iter__
    raise e
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/fastprogress/fastprogress.py", line 41, in __iter__
    for i,o in enumerate(self.gen):
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 628, in __next__
    data = self._next_data()
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1333, in _next_data
    return self._process_data(data)
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1359, in _process_data
    data.reraise()
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/torch/_utils.py", line 543, in reraise
    raise exception
FileNotFoundError: Caught FileNotFoundError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop
    data = fetcher.fetch(index)
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/torch/utils/data/_utils/fetch.py", line 58, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/torch/utils/data/_utils/fetch.py", line 58, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/media/cvpr/CM_22/doctr/doctr/datasets/datasets/base.py", line 48, in __getitem__
    img, target = self._read_sample(index)
  File "/media/cvpr/CM_22/doctr/doctr/datasets/datasets/pytorch.py", line 37, in _read_sample
    else read_img_as_tensor(os.path.join(self.root, img_name), dtype=torch.float32)
  File "/media/cvpr/CM_22/doctr/doctr/io/image/pytorch.py", line 52, in read_img_as_tensor
    pil_img = Image.open(img_path, mode="r").convert("RGB")
  File "/home/cvpr/anaconda3/envs/pytesseract/lib/python3.9/site-packages/PIL/Image.py", line 2912, in open
    fp = builtins.open(filename, "rb")
FileNotFoundError: [Errno 2] No such file or directory: '/media/cvpr/CM_22/mjsynth/mnt/ramdisk/max/90kDICT32px/긴장_1227682.jpg'

解决方案 #

问题核心是你直接把其他数据集的data元素追加到MJSynth实例中，但MJSynth的读取逻辑会固定拼接自身的root路径（MJSynth图片目录），而自定义数据集和SynthText的图片并不在该目录下，因此报错。

方法1：用ConcatDataset合并数据集（推荐） #

PyTorch的ConcatDataset可以直接合并多个独立的Dataset实例，每个实例会用自己的路径逻辑读取图片，不会互相干扰：

from torch.utils.data import ConcatDataset

# 保留原有的三个数据集初始化代码
custom_train_set = RecognitionDataset(
    parts[0].joinpath("images"),
    parts[0].joinpath("labels.json"),
    img_transforms=Compose(
        [
            T.Resize((args.input_size, 4 * args.input_size), preserve_aspect_ratio=True),
            T.RandomApply(T.ColorInversion(), 0.1),
            ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.02),
        ]
    ),
)
if len(parts) > 1:
    for subfolder in parts[1:]:
        custom_train_set.merge_dataset(
            RecognitionDataset(subfolder.joinpath("images"), subfolder.joinpath("labels.json"))
        )

train_set_mjsynth = MJSynth(
    train=True,
    img_folder='/media/cvpr/CM_22/mjsynth/mnt/ramdisk/max/90kDICT32px',
    label_path='/media/cvpr/CM_22/mjsynth/mnt/ramdisk/max/90kDICT32px/imlist.txt',
    img_transforms=T.Resize((args.input_size, 4 * args.input_size), preserve_aspect_ratio=True),
)

train_set_synthtext = SynthText(
    train=True,
    recognition_task=True,
    download=True,
    img_transforms=T.Resize((args.input_size, 4 * args.input_size), preserve_aspect_ratio=True),
)

# 合并三个数据集
train_set = ConcatDataset([train_set_mjsynth, train_set_synthtext, custom_train_set])

方法2：修改数据集存储逻辑（不推荐） #

如果一定要用extend方式，需要让所有数据集的data存储完整绝对路径而非仅图片名称，同时修改_read_sample方法直接使用该路径读取图片。但这种方式需要改动底层数据集代码，复杂度高，优先选方法1。

额外注意点 #

• 确保三个数据集的__getitem__返回格式一致（图片张量+目标标签），ConcatDataset才能正常工作。
• 自定义数据集有增强操作，MJSynth和SynthText没有，若需要统一增强，可将增强逻辑移到DataLoader的collate_fn，或统一设置到所有数据集的img_transforms中。

内容的提问来源于stack exchange，提问作者Khawar Islam