torch.cat error in evaluation

[lianqiann]

I think the usage of torch.cat is wrong. And this would cause a runtime error

https://github.com/UX-Decoder/DINOv/blob/53bf20d5cfdbb86fa35141a1cff432d4923599f2/dinov/architectures/dinov.py#L1082

When running eval script for COCO2017 python train_net.py --eval_only --resume --eval_get_content_features --num-gpus 8 --config-file /path/to/configs COCO.TEST.BATCH_SIZE_TOTAL=8 MODEL.WEIGHTS=/path/to/weights OUTPUT_DIR=/path/to/outputs

This would cause error that

all_rand_shape = torch.cat([t['rand_shape'] for t in new_targets], 0)
RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 800 but got size 896 for tensor number 1 in the list.

The tensors causing such error are in following shapes

torch.Size([30, 800, 1216])
torch.Size([2, 896, 800])
torch.Size([20, 800, 1088])
torch.Size([8, 1088, 800])

This issue will cause many other errors in many places, so fixing this alone cannot make the eval script run. Could anyone check on this? Thank you very much!

[lianqiann]

i think the error is from that the targets size were not unified for inputs with different sizes. This issue only happens in non-training mode.

[FengLi-ust]

Yes, in training we resize all images to a fixed size. For inference, using batch size 1 should work, otherwise you need to pad the images into the same size for batched inference.