[Fix] Support amp (pytorch >= 1.6.0) on DCN and DCNv2/ Add unit tests on DCN/DCNv2 amp #1029
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
|
Codecov Report
@@ Coverage Diff @@
## master #1029 +/- ##
==========================================
- Coverage 65.35% 65.34% -0.01%
==========================================
Files 155 155
Lines 9927 9930 +3
Branches 1806 1806
==========================================
+ Hits 6488 6489 +1
- Misses 3105 3107 +2
Partials 334 334
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
AronLin
changed the title
xvjiarui
reviewed
May 20, 2021
mmcv/ops/deform_conv.py
Outdated
Comment on lines
73
to
76
| # The flag for whether to use fp16 (pytorch < 1.6.0) or | ||
| # amp (pytorch >= 1.6.0) is the type of "offset", we | ||
| # cast weight and input to temporarily support fp16 and | ||
| # amp whatever the pytorch version is. |
There was a problem hiding this comment.
We may state why we are doing this, e.g amp won't cast input for DCN correctly.
xvjiarui
approved these changes
May 21, 2021
ZwwWayne
reviewed
May 23, 2021
ZwwWayne
approved these changes
May 23, 2021
ZwwWayne
approved these changes
May 23, 2021
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation
mmcv 1.3.2 supports automatic mixed precision with pytorch >= 1.6.0 but the CUDA ops don't support it now, resulting in some runtime errors like #1004 and #1028.
#1014 fixes the bug in #1004. However, when the type of input data for DCN is float32, the bug still exists. The same bug also exits in DCNv2.
Here is I'd like to do:
Modification
The offset in deform_conv is computed by a Conv2d which supports amp, so the type of offset is the flag for whether we use fp16 or amp.
With amp (pytorch >= 1.6.0), the model type will not be set manually. The input type might be float32 and float16; the unit tests should include both cases.
BC-breaking (Optional)
No.