Class detection works for huggingface checkpoints #1800
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
mattdangerw
force-pushed
the
from-preset-improvements
branch
3 times, most recently
from
893fad0
to
c2fe6e8
Compare
This was a tricky one to fix that involved some large refactoring to our preset loading routines. Originally the intent was that `from_preset()` was a easily readable bunch of lower-level Keras calls. With the arrival of transformers conversions, and soon timm conversions, I think that goal is no longer super realistic. Instead I added a loader interface, with default implementations off `load_task` and `load_preprocessor`. Every format we support directly converting from has to support at a minimum... - Detecting the backbone class. - Loading the backbone class. One consequence of this work is that every class with a `from_preset` constructor needs to reference the `backbone_cls` they match with. I think this will be a more stable way to handle our "auto class" like functionality as we venture further towards multi-modal models
mattdangerw
force-pushed
the
from-preset-improvements
branch
from
c2fe6e8
to
3e7f29c
Compare
SamanehSaadat
approved these changes
Aug 27, 2024
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Outdated suggestions cannot be applied.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.
Fixes #1798. You can now do
keras_nlp.models.Backbone("hf://google-bert/bert-base-uncased")with a safetensors checkpoint, are we will find the correct architecture to instantiate. This has always worked for Keras-style checkpoints, but not safetensor ones.This was a tricky one to fix that involved some large refactoring to our preset loading routine.
Originally the intent was that
from_preset()was a easily readable bunch of lower-level Keras calls. With the arrival of transformers conversions, and soon timm conversions, I think that goal is no longer super realistic. Instead I added a loader interface, with default implementations offload_taskandload_preprocessor. Every format we support directly converting from has to support at a minimum...One consequence of this work is that every class with a
from_presetconstructor needs to reference thebackbone_clsthey match with. I think this will be a more stable way to handle our "auto class" like functionality as we venture further towards multi-modal models