Wrong shape order in TokenInitailizer?

[Birdylx]

@ahatamiz Hi, thanks for you work, but it seems the shape over here is in wrong order?

[ahatamiz]

Hi @Birdylx thank you for your meticulous attention to the details 👍

#8 should address this issue.

[Birdylx]

@ahatamiz, I wonder if you're using square image to train objection and segmentation model, if not, seems this wrong tokenization doesn't hurt performance.

[ahatamiz]

Thanks again @Birdylx for the feedback. Fortunately, this only happened during refactoring for the newly added faster_vit_any_res model and did not impact any of the experiments. And also as indicated in the paper, experiments are done on non-square image size for segmentation, detection.

I also agree with you about the performance. It should not still impact the performance, but we also made sure that shape order is correct in TokenInitializer.