Closed flyfaerss closed 1 year ago
Hi, thanks for your interest in our work!
The codes you mentioned are what we simply adopt from the original repo of DETR:
https://github.com/facebookresearch/detr/blob/8a144f83a287f4d3fece4acdf073f387c5af387d/models/segmentation.py#L59
In our experiments, we set the batch size as 1 (per GPU), so there isn't any problem.
We haven't tried with a larger batch_size.
But I think that you might be right, the permute/transpose function is more appropriate for the case
In line 323 and 329 on psgtr_head.py file, it is confused to use function 'view' to change shape (100, 3, ...) to (3, 100, ...). Do you mean to convert query unit into batch unit? I don't think the 'view' function can achieve this goal. Why isn't it suitable for permute or transpose?