-
Notifications
You must be signed in to change notification settings - Fork 3.5k
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
[Quantization] Make calibration faster and more memory usage friendly #4589
Merged
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
vinx13
approved these changes
Jan 3, 2020
Thanks @masahi this is merged |
alexwong
pushed a commit
to alexwong/tvm
that referenced
this pull request
Feb 26, 2020
…apache#4589) * Use memory efficient calibrate * Fixed indexing * add cpp kl stub * ported KL cpp from mxnet * Fixed std::distance arguments order * remove python implementation * fix lint and indent * fix indent * refactoring * fix lint * fix for i386
alexwong
pushed a commit
to alexwong/tvm
that referenced
this pull request
Feb 28, 2020
…apache#4589) * Use memory efficient calibrate * Fixed indexing * add cpp kl stub * ported KL cpp from mxnet * Fixed std::distance arguments order * remove python implementation * fix lint and indent * fix indent * refactoring * fix lint * fix for i386
zhiics
pushed a commit
to neo-ai/tvm
that referenced
this pull request
Mar 2, 2020
…apache#4589) * Use memory efficient calibrate * Fixed indexing * add cpp kl stub * ported KL cpp from mxnet * Fixed std::distance arguments order * remove python implementation * fix lint and indent * fix indent * refactoring * fix lint * fix for i386
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Outdated suggestions cannot be applied.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.
This PR improves the performance (not accuracy) of KL based calibration in two ways:
Current implementation stores entire samples for all layers in one go, this quickly becomes intractable in terms of memory usage for image segmentation tasks, where there are more than a hundred of intermediate outputs and high-res inputs are common. I added "calibrate_chunk_by" parameter to qconfig, which enables chunk-by-chunk, interleaved profile generation and scale calculation. This adds some redundant computation, but it is a worthwhile trade off. In practice, using the cuda target for profile generation, I don't find noticeable slowdown. The default behavior is to use the number of intermediate outputs as chunk size, so there will be only one chunk and performance is the same as existing implementation.
Port KL div minimization to C++. MXNet has C++ implementation now, so I replaced python one we have with that one. This turned out a big win, scale calculation is now 10-20x faster. Below is a log from running test_calibrate_chunk() that I added, showing elapsed seconds and found scales for python and cpp respectively.
please review @vinx13 @ZihengJiang @tmoreau89