Add MobileNetV5 to KerasHub #2399
Conversation
harshaljanjani
added
the
WIP
harshaljanjani
added
kokoro:force-run
|
Let's start the review rounds now that the model numerics match within a 1e-5 tolerance and the design integrates with Gemma 3n!! |
|
/gemini review |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces MobileNetV5 to KerasHub, following the established modular structure with a backbone, classifier, preprocessor, tests, and a checkpoint conversion script. The implementation is comprehensive and well-organized. My primary feedback revolves around ensuring model serializability. Several new custom layers are missing the get_config method, which is critical for saving and loading models. I have detailed these instances in my comments. Additionally, I've identified a couple of latent bugs in the MobileNetV5Builder concerning the handling of the 'ir' block type, which, while not affecting current presets, should be addressed for future robustness. Overall, this is a solid contribution, and addressing the serialization issues is the main priority.
There was a problem hiding this comment.
Thank you for the contribution, I have added my comment.
Can you add the test smallest preset test, you can mark as skip test for now.
Also, add on the fly conversion from hf here https://github.com/keras-team/keras-hub/tree/master/keras_hub/src/utils/timm
|
@sachinprasadhs Resolved all comments; thanks!
Since only one preset is currently exposed for MobileNetV5, are these two additions really needed? It might make sense to keep them if other presets are later added by TIMM / PyTorch Image Models, but for now, I’m not too sure; since all presets == the smallest preset, and there’s only one preset to load from HF, which we’re already providing with our Kaggle preset. Would love your suggestions! |
Usually for the models ported from HF, we require the on the fly conversion as well so that the model can be loaded from Kaggle as well as HF. |
Makes sense, will cover the same! |
|
None of the test failures are related to MobileNetV5. TensorFlow (GPU): GPT-2 
JAX (GPU): Timeout 
|
| arch_def = [["er_r1_k3_s2_e4_c24"], ["uir_r2_k5_s2_e6_c48"]] | ||
| block_args = decode_arch_def(arch_def) | ||
| backbone = keras_hub.models.MobileNetV5Backbone(block_args=block_args) | ||
| model = keras_hub.models.MobileNetV5ImageClassifier(backbone, 1000) |
There was a problem hiding this comment.
This 1000 classes will be of random weights right?, if HF weights does not have any classifier head and the backbone also is of random weights here and not from the preset.
There was a problem hiding this comment.
Yeah, the examples aren’t demonstrating preset loading; they’re just showing a forward pass of a randomly initialized instance, which also proves the point. But this review made me realize that I haven’t updated the example for the image classifier yet. I’ll do the needful.
Also, for preset loading, I think I’ll add an example that demonstrates running predictions using a preset as well.
Edit: Done.
The same still holds true. |
The current design integrates well within the Gemma 3n scope. Ready for reviews!
Numerics Matching (Precision: 1e-5)
cc: @divyashreepathihalli @mattdangerw
Colab Notebook
Numerics Validation Reproducibility Notebook
Closes #2401.