Stack error on using mini-batch size > 1

Describe the bug On running the below code:

cl_strategy = EWC(
    model_ft, optimizer_ft, criterion, ewc_lambda=5,
    train_mb_size=32, train_epochs=4, eval_mb_size=32
)

We get the below-mentioned error. This works fine for train_mb_size=1 and eval_mb_size=1 but gives error for any other value.

/usr/local/lib/python3.7/dist-packages/torch/utils/data/_utils/collate.py in default_collate(batch)
     53             storage = elem.storage()._new_shared(numel)
     54             out = elem.new(storage)
---> 55         return torch.stack(batch, 0, out=out)
     56     elif elem_type.__module__ == 'numpy' and elem_type.__name__ != 'str_' \
     57             and elem_type.__name__ != 'string_':

RuntimeError: stack expects each tensor to be equal size, but got [3, 341, 500] at entry 0 and [3, 313, 500] at entry 1

To Reproduce

cl_strategy = EWC(
    model_ft, optimizer_ft, criterion, ewc_lambda=5,
    train_mb_size=32, train_epochs=4, eval_mb_size=32
)

Expected behavior It was supposed to train with the specified mini-batch size.

[AndreaCossu]

I inspected the CUB200 dataset and the problem is related to the fact that each pattern is a 3-channel image with different height and width. So, each pattern tensors has shape (3, H, W). The dataloader tries to build the minibatch with stack and fails since dimensions are different.
I think this could be fixed by forcing a center crop of same dimensions on all patterns or by padding. However, I don’t know if there are standard practices to work with this dataset. @lrzpellegrini , @vlomonaco do you have any hints on this?

[lrzpellegrini]

Ok, I’ll try to check if a sproper implementation of the dataset exists in the wild 😉