You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2021/02/20 00:25:07 UTC
[GitHub] [tvm] trevor-m opened a new pull request #7483: [TOPI] Fix cuda nms handling of additional per box features
trevor-m opened a new pull request #7483:
URL: https://github.com/apache/tvm/pull/7483
For NMS, boxes typically have 5 or 6 features which are the 4 box coordinates, per-box scores, and sometimes per-box classes. However, the boxes are also allowed to have any amount of additional features. We didn't have any unit tests for that situation, so have added one. After recent changes to CUDA nms implementation, those additional features were not being copied around anymore.
Additional features per box: https://mxnet.incubator.apache.org/versions/1.7.0/api/python/docs/api/symbol/contrib/index.html#mxnet.symbol.contrib.box_nms
> By default, a box is [id, score, xmin, ymin, xmax, ymax, …], additional elements are allowed.
@masahi @mbrookhart @anijain2305
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] masahi merged pull request #7483: [TOPI] Fix cuda nms handling of additional per box features
Posted by GitBox <gi...@apache.org>.
masahi merged pull request #7483:
URL: https://github.com/apache/tvm/pull/7483
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] masahi commented on pull request #7483: [TOPI] Fix cuda nms handling of additional per box features
Posted by GitBox <gi...@apache.org>.
masahi commented on pull request #7483:
URL: https://github.com/apache/tvm/pull/7483#issuecomment-783678203
Thanks @trevor-m @mbrookhart
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] trevor-m commented on pull request #7483: [TOPI] Fix cuda nms handling of additional per box features
Posted by GitBox <gi...@apache.org>.
trevor-m commented on pull request #7483:
URL: https://github.com/apache/tvm/pull/7483#issuecomment-783638557
> @trevor-m Thanks, btw is the PyTorch 4500 boxes number coming from a real workload? It seems really slow. If it is multiclass NMS, I suggest using graph rewrite I added for it (see `pytorch/test_object_detection.py`)
No it's not real, I just edited the box count for the Pytorch unit test `test_forward_nms`
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] masahi commented on pull request #7483: [TOPI] Fix cuda nms handling of additional per box features
Posted by GitBox <gi...@apache.org>.
masahi commented on pull request #7483:
URL: https://github.com/apache/tvm/pull/7483#issuecomment-783620868
@trevor-m Thanks, btw is the PyTorch 4500 boxes number coming from a real workload? It seems really slow. If it is multiclass NMS, I suggest using graph rewrite I added for it (see `pytorch/test_object_detection.py`)
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] trevor-m commented on pull request #7483: [TOPI] Fix cuda nms handling of additional per box features
Posted by GitBox <gi...@apache.org>.
trevor-m commented on pull request #7483:
URL: https://github.com/apache/tvm/pull/7483#issuecomment-783606347
@masahi Thank you for reviewing! The performance seems to be the same with same random variance.
| Frontend | #Boxes | Without this PR | With this PR |
|---|---|---|---|
| Pytorch | 64 | 0.05 ms | 0.04 ms |
| Pytorch | 1024 | 1.10 ms | 1.09 ms |
| Pytorch | 4500 | 44.73 ms | 44.59 ms |
| TensorFlow | 64 | 0.05 ms | 0.05 ms |
| TensorFlow | 1024 | 0.11 ms | 0.10 ms |
| TensorFlow | 67500 | 4.91 ms | 4.95 ms |
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] masahi commented on pull request #7483: [TOPI] Fix cuda nms handling of additional per box features
Posted by GitBox <gi...@apache.org>.
masahi commented on pull request #7483:
URL: https://github.com/apache/tvm/pull/7483#issuecomment-782510416
please verify that there is no perf regression for normal cases (PyTorch, TF). I now understand why MXNet NMS expects weird packed inputs lol
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org