NielsRogge

@ArthurZucker thanks, however could you clarify? Your code snippet seems to use id = 64.000 for the "" token, whereas the "\" token needs to have that index. Do you...

Pinging @qubvel here

I'll mark this issue as a "good second issue" for someone who could look into the details of the run_mae.py script. For now I recommend to leverage the original MAE...

@CristoJV you can follow this guide: https://github.com/facebookresearch/mae/blob/main/PRETRAIN.md. The script supports a `--resume` flag, hence you can run the script with the following flag in order to further pre-train the model:...

Hi @EduardoPach do you need any help in finishing this PR? Really great to see you're leveraging `Copied from` for the text encoder and all parts taken from Deformable DETR....

Thanks @amyeroberts for the review! Together with @EduardoPach, we've addressed all comments. CI is green, one unrelated test is failing. See also my comments above regarding why we wouldn't use...

cc @younesbelkada

cc @younesbelkada I think the thread above is correct, either: * `input_ids` and `labels` are the same but then one needs to shift the `logits` when computing the loss. Examples...

@ArthurZucker the tokenizer is the only thing left to make all tests green. The PR is ready other than that. The only issue that is remaining are the sentinel tokens...

Update: we're down to 2 failing tests: ``` FAILED tests/models/udop/test_processor_udop.py::UdopProcessorTest::test_save_load_pretrained_default - AssertionError: {'▁backing': 16057, '▁Brunswick': 29980, 'S[629176 chars]7501} != {'': 0, '': 1, '': 2, '▁': 3,[624686 chars]4401} FAILED tests/models/udop/test_tokenization_udop.py::UdopTokenizationTest::test_save_slow_from_fast_and_reload_fast...