Support Loading Quantized Models with from_preset()

[JyotinderSingh]

Description of the change

This change resolves an issue with loading quantized models from presets. Previously, the model's serialized DTypePolicyMap was not correctly passed to the backbone during loading, which caused failures during initialization of quantized layers.

The fix introduces a new _resolve_dtype utility function that determines the correct dtype for the model based on the following rules:

1. User-specified dtype: If a user explicitly provides a dtype in the from_preset call (e.g., from_preset("bert_tiny_en_uncased", num_classes=2, dtype="float32")), that value is used.

2. Float type casting: If no user dtype is provided and the saved dtype is a floating-point type (e.g., "float32"), the model will be loaded using the current Keras default dtype policy. This allows for safe casting between different floating-point precisions.

3. DTypePolicyMap: If no user dtype is provided and the saved dtype is a complex object (like a DTypePolicyMap for quantization), the saved type is used as is. This ensures that quantization configurations are preserved during loading.

Colab Notebook

1. Failing Case: This notebook demonstrates the original bug where loading the preset fails.
2. Successful Case: This notebook shows the corrected behavior where the quantized preset loads successfully with this change.

Checklist

• I have added all the necessary unit tests for my change.
• I have verified that my change does not break existing code and works with all backends (TensorFlow, JAX, and PyTorch).
• My PR is based on the latest changes of the main branch (if unsure, rebase the code).
• I have followed the Keras Hub Model contribution guidelines in making these changes.
• I have followed the Keras Hub API design guidelines in making these changes.
• I have signed the Contributor License Agreement.

[mattdangerw]

Thanks!

[JyotinderSingh]

Resolved comments

[divyashreepathihalli]

/gemini review

[gemini-code-assist]

Code Review

This pull request effectively addresses an issue with loading quantized models from presets by introducing a _resolve_dtype utility function and ensuring dtype policies are correctly serialized. The changes are logical and well-tested. I have a couple of minor suggestions to fix a test assertion message and improve docstring formatting to align with the style guide.

[mattdangerw]

lgtm! just a couple nits

[mattdangerw]

Thanks!