Open CarlosGomes98 opened 1 month ago
That's strange. When using the test tests/test_backbones.py::test_vit_models_non_divisible_input
(from the branch associated to this issue) I got:
> raise EinopsError(message + "\n {}".format(e))
E einops.EinopsError: Error while processing rearrange-reduction pattern "b c (t tub) (h p) (w q) -> b (t h w) (tub p q c)".
E Input tensor shape: torch.Size([1, 6, 4, 220, 230]). Additional info: {'tub': 1, 'p': 16, 'q': 16}.
E Shape mismatch, can't divide axis of length 220 in chunks of 16
Isn't that the expected behaviour ?
Describe the issue With prithvi_vit, when the input spatial dimensions are not divisible by the patch size, part of the input is ignored. To Reproduce (optional, but appreciated) Steps to reproduce the behavior:
Expected behavior (optional) Either we should pad the input to a size divisible by the patch size, or throw an error