bug: model saving and loading does not work with Keras 3.x model format #4921
Description
Describe the bug
Since Keras 3 (released Novembre 28, 2023), Keras model saving and loading API changed and now requires a filepath for the model ending iwith .keras as opposed to a folder. The .keras format is simply a compressed folder.
Error
As such, the bentoml.keras.save_model and bentoml.keras.load_model API are not working as expected anymore, since they try to save and load from a folder path.
ValueError: Invalid filepath extension for saving. Please add either a `.keras` extension for the native Keras format (recommended) or a `.h5` extension. Use `model.export(filepath)` if you want to export a SavedModel for use with TFLite/TFServing/etc. Received: filepath=/tmp/tmp394vyky1bentoml_model_my_keras_model/.
As far as I can tell, BentoML's Keras framework layer API have been broken since Keras 3 release.
Fix
Starting with Keras 3.4.0 (released June 25, 2024), Keras allows to pass a new zipped boolean parameter which allows the older behaviour (see keras-team/keras@e0eaac4).
This can be used to restore the compatibility with BentoML when saving and loading Keras models.
To reproduce
Save a keras model with BentoML in its model store, using a recent version of Keras (3.x) or Tensorflow.
# Save the model using BentoML to its model store
# https://docs.bentoml.com/en/latest/reference/frameworks/keras.html#bentoml.keras.save_model
bentoml.keras.save_model(
"my_keras_model",
model,
include_optimizer=True,
custom_objects={
"preprocess": preprocess,
"postprocess": postprocess,
},
)
Traceback:
Traceback (most recent call last):
File "/home/remy/my_project/src/train.py", line 134, in <module>
main()
File "/home/remy/my_project/src/train.py", line 111, in main
bentoml.keras.save_model(
File "/home/remy/my_project/.venv/lib/python3.11/site-packages/bentoml/_internal/frameworks/keras.py", line 272, in save_model
model.save(bento_model.path, include_optimizer=include_optimizer, **kwargs)
File "/home/remy/my_project/.venv/lib/python3.11/site-packages/keras/src/utils/traceback_utils.py", line 122, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/home/remy/my_project/.venv/lib/python3.11/site-packages/keras/src/saving/saving_api.py", line 114, in save_model
raise ValueError(
ValueError: Invalid filepath extension for saving. Please add either a `.keras` extension for the native Keras format (recommended) or a `.h5` extension. Use `model.export(filepath)` if you want to export a SavedModel for use with TFLite/TFServing/etc. Received: filepath=/tmp/tmp394vyky1bentoml_model_my_keras_model/.
Expected behavior
Model is correctly saved with bentoml.keras.save_model.
Environment
bentoml: 1.3.2
keras: 3.5.0 (transitively through tensorflow 2.17)
python: 3.11
os: Linux