YosuaMichael

@TeodorPoncu Oh ya, also please resolve the conflict on the test files

NOTE: I also include the flaky autocast on the commit: Fix increase score_thresh for FCOS

> Some bug with relative position bias . Running locally I can trace error here. My suspicion is that we are doing something wrong in > > ``` > def...

Also looking at the failing test, seems like it has something to do with different result on the expected file

> Also @YosuaMichael Did you check with Omnivore team about the minor differences we noticed, apart from patch_size which is now [2, 4, 4]. I think the weight provided by...

Hmm... the differences in expected file might due to different platform you use to generate it. Let me try to generate the expected file and see if I got the...

Thanks @oke-aditya for the ported weight. I [modify](https://github.com/YosuaMichael/vision/blob/swin3d_validation/torchvision/models/video/swin_transformer.py#L508) the code from this branch a bit by adding the weight and transforms to mimic the original paper. Then I ran the...

Looks good @oke-aditya , I agree with the naming `KINETICS_IMAGENET22K_V1`! For `min_size, temporal_size, and num_params` it is as @datumbox explained. I think swin transformer 3d dont really have `min_size` and...

cc @jdsgomes

I think this case is implicitly guarded on https://github.com/pytorch/vision/blob/d5bd8b728f14c33b339fc45c90ca39be339bce3f/references/classification/train.py#L87 since `len(data_loader.dataset) != num_processed_samples` shouldn't be true on non-distributed setting. Do you get the error during non-distributed training @pmeier ?