[keras 3.4.1, tf 2.16.2] loading & saving models impossible

[tobwen]

issues

There are issues with loading and saving models, which are obviously related to deprecated APIs of Tensorflow and Keras.

examples

saving

model.save('test.keras')

This assert pops up and asks me to save as ".keras" (setting a save-format is deprecated).

uncertainty-wizard/uncertainty_wizard/models/_stochastic/_abstract_stochastic.py

Lines 203 to 211 in ec6600d

	assert (
	not filepath.lower().endswith("h5")
	and not filepath.lower().endswith("hdf5")
	and not filepath.lower().endswith(".keras")
	and (save_format is None or not save_format.lower() == "h5")
	), (
	"Uncertainty Wizard does not support the deprecated h5 format to save models."
	"Change the file ending or the save_format parameter to save using the better tf `SavedModel` format."
	)

When removing this, I'm getting

ValueError: The following argument(s) are not supported: ['signatures', 'options']

uncertainty-wizard/uncertainty_wizard/models/_stochastic/_abstract_stochastic.py

Lines 218 to 219 in ec6600d

	signatures=signatures,
	options=options,

When removing these parameters, saving works... but I can't reload the model (JSON is shortened)

{"module": "keras", "class_name": "Sequential", "config": {"name": "sequential", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 140376256590384}, "layers": [{"module": "keras.layers", "class_name": "InputLayer", 

loading

Although I'm on tf v2.16.2, Load-Options don't seem to be supported, so neither uwiz.models.load_model() nor StochasticSequential.load_model() nor native models.load_model() work:

uncertainty-wizard/uncertainty_wizard/models/_load_model.py

Line 72 in ec6600d

path, custom_objects=custom_objects, compile=compile, options=options

Gets me:

Exception encountered: <class 'uncertainty_wizard.models.stochastic_utils.layers.UwizBernoulliDropout'> could not be deserialized properly. Please ensure that components that are Python object instances (layers, models, etc.) returned by `get_config()` are explicitly deserialized in the model's `from_config()` method.

config={'module': 'uncertainty_wizard.models.stochastic_utils.layers', 'class_name': 'UwizBernoulliDropout', 'config': {'name': 'UwBernoulliDropout', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None, 'shared_object_id': 140376256590384}, 'rate': 0.5, 'seed': None, 'noise_shape': None}, 'registered_name': 'Custom>UwizBernoulliDropout', 'build_config': {'input_shape': [None, 24, 24, 64]}}.

Exception encountered: Error when deserializing class 'UwizBernoulliDropout' using config={'name': 'UwBernoulliDropout', 'trainable': True, 'dtype': 'float32', 'rate': 0.5, 'seed': None, 'noise_shape': None}.

Exception encountered: UwizBernoulliDropout.__init__() missing 1 required positional argument: 'stochastic_mode'

Actually, this is lacking stochastic_mode:

uncertainty-wizard/uncertainty_wizard/models/stochastic_utils/layers.py

Lines 108 to 112 in ec6600d

	# docstr-coverage:inherited
	def get_config(self):
	config = super(UwizBernoulliDropout, self).get_config()
	config["name"] = "UwBernoulliDropout"
	return config

Let's try:

    def get_config(self):
        config = super(UwizBernoulliDropout, self).get_config()
        config["name"] = "UwBernoulliDropout"
        config["stochastic_mode"] = self.stochastic_mode # new
        return config

Nope, didn't fix it. Sorry, I'm stuck.

environment

• keras: v3.4.1
• tensorflow: v2.16.2