Creates standard for PreTrained behavior #2360
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
This changes the standard for DJL behavior with preTrained blocks. As of now, they should also start out with frozen parameters. This has been applied to the embeddings. It was previously applied only to PyTorch, but as of now applies to all models. However, I did leave a carveout for models. It adds a boolean "wasLoaded" so that if you load a model and then create a Trainer directly from it, it will not be frozen. If you load a model and then append some new layers to it (as we have several examples of), then it will need to be unfrozen.
Codecov ReportBase: 72.08% // Head: 74.37% // Increases project coverage by
📣 This organization is not using Codecov’s GitHub App Integration. We recommend you install it so Codecov can continue to function properly for your repositories. Learn more Additional details and impacted files@@ Coverage Diff @@
## master #2360 +/- ##
============================================
+ Coverage 72.08% 74.37% +2.28%
- Complexity 5126 6817 +1691
============================================
Files 473 670 +197
Lines 21970 29599 +7629
Branches 2351 3073 +722
============================================
+ Hits 15838 22013 +6175
- Misses 4925 6086 +1161
- Partials 1207 1500 +293
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. ☔ View full report at Codecov. |
frankfliu
approved these changes
Feb 2, 2023
zachgk
added a commit
to zachgk/djl
that referenced
this pull request
Feb 14, 2023
In deepjavalibrary#2360, the behavior of using pre-trained models was to freeze parameters. However freezing the parameters on MXNet seems to cause a significant performance regression for training. This removes those changes for a temporary workaround until a deeper investigation can take place.
frankfliu
added a commit
that referenced
this pull request
Feb 14, 2023
* Remove performance issues from freezing MXNet In #2360, the behavior of using pre-trained models was to freeze parameters. However freezing the parameters on MXNet seems to cause a significant performance regression for training. This removes those changes for a temporary workaround until a deeper investigation can take place. Co-authored-by: Frank Liu <frankfliu2000@gmail.com>
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 changes the standard for DJL behavior with preTrained blocks. As of now, they should also start out with frozen parameters. This has been applied to the embeddings.
It was previously applied only to PyTorch, but as of now applies to all models. However, I did leave a carveout for models. It adds a boolean "wasLoaded" so that if you load a model and then create a Trainer directly from it, it will not be frozen. If you load a model and then append some new layers to it (as we have several examples of), then it will need to be unfrozen to retrain.
fixes #2351