Potentially wrong inherence in lavis.datasets.datasets.base_dataset

[hexiaoxiao-cs]

When I try the run_scripts/blip2/pretrain_stage1.sh, an error happened in the dataloader:

AttributeError('Caught AttributeError in DataLoader worker process 0.\nOriginal Traceback (most recent call last):\n File "//.conda/envs/lavis/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 308, in _worker_loop\n data = fetcher.fetch(index)\n File "//.conda/envs/lavis/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 54, in fetch\n return self.collate_fn(data)\n File "", line 2, in collater\n File "/***/.conda/envs/lavis/lib/python3.8/contextlib.py", line 111, in enter\n del self.args, self.kwds, self.func\nAttributeError: args\n')

After debugging, the problems comes from the class ConcatDataset in lavis/datasetes/datasets/base_dataset.

I've changed the line 13 to from torch.utils.data import Dataset, ConcatDataset as _ConcatDataset and line 74 to class ConcatDataset(_ConcatDataset):, and the problem goes away.

Here is the cause of the problem (what I believe, if not please correct me). The ConcatDataset should inherit from torch.dataset.ConcatDataset and it does not initialize correctly. Then, when the dataloader tries to get a batch of data, the dataloader did not find a getitems function under the ConcatDataset class (since it is not inherited correctly), thus raise the exception.