-
Notifications
You must be signed in to change notification settings - Fork 1.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[WIP] Accelerate training by replacing DataContainer object scatter #1236
Conversation
Codecov Report
@@ Coverage Diff @@
## master #1236 +/- ##
==========================================
+ Coverage 68.27% 68.29% +0.02%
==========================================
Files 160 160
Lines 10599 10597 -2
Branches 1937 1937
==========================================
+ Hits 7236 7237 +1
+ Misses 2979 2976 -3
Partials 384 384
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
What's ur using case? Does that only work for |
His behavior is exactly the same as before, just replaced with the scatter of Pytorch itself. |
I tried this PR in mmdetection3D on PointPillar: Environment: original implementations:
using this PR:
The training speed seems faster and accuracy seems higher after using this PR. |
We found that there is no need to modify it temporarily, so it is closed. If there are new developments in the follow-up, it will start again. |
Motivation
During the YOLOX reproduction process, we found that the
scatter
process of DataContainer will significantly increase the training time. The practice has shown that replacing the custom scatter can reduce the training time by about half.Modification
Replace the custom scatter of the DataContainer object with the scatter of PyTorch.
BC-breaking
None
Use cases
I only tested MMDetection, the other frameworks did not test.
Note
I need to experiment and compare the training and inference speed.