Skip to content

Commit

Permalink
[Feature] dev-1.x change np.transpose to torch.permute for speed up (o…
Browse files Browse the repository at this point in the history
…pen-mmlab#2277)

* change np.transpose to torch.permute

* add comments
  • Loading branch information
sunjiahao1999 authored and xizaoqu committed Mar 15, 2023
1 parent 6341a47 commit ad1c2d5
Showing 1 changed file with 8 additions and 4 deletions.
12 changes: 8 additions & 4 deletions mmdet3d/datasets/transforms/formating.py
Original file line number Diff line number Diff line change
Expand Up @@ -147,15 +147,19 @@ def pack_single_results(self, results: dict) -> dict:
if 'img' in results:
if isinstance(results['img'], list):
# process multiple imgs in single frame
imgs = [img.transpose(2, 0, 1) for img in results['img']]
imgs = np.ascontiguousarray(np.stack(imgs, axis=0))
imgs = [to_tensor(img) for img in results['img']]
imgs = torch.stack(
imgs, dim=0).permute(0, 3, 1, 2).contiguous()
results['img'] = to_tensor(imgs)
else:
img = results['img']
if len(img.shape) < 3:
img = np.expand_dims(img, -1)
results['img'] = to_tensor(
np.ascontiguousarray(img.transpose(2, 0, 1)))
# To improve the computational speed by by 3-5 times, apply:
# `torch.permute()` rather than `np.transpose()`.
# Refer to https://github.com/open-mmlab/mmdetection/pull/9533
# for more details
results['img'] = to_tensor(img).permute(2, 0, 1).contiguous()

for key in [
'proposals', 'gt_bboxes', 'gt_bboxes_ignore', 'gt_labels',
Expand Down

0 comments on commit ad1c2d5

Please # to comment.