Loading a model created and trained in Keras 2. in new Keras 3.7 env for inference

[demybug123]

I am working on a computer vision project and I have this model that is trained in Keras 2. and saved as an .h5 file.
In the 'old' Keras 2. environment, I can load the model with custom and run predict just fine.

Here's the code snippet I use for loading model and custom eswish activation function.

from tensorflow.keras.backend import set_learning_phase
from tensorflow.keras.models import load_model

class eswish(Activation):
    """
    Custom Swish activation function for Keras.
    """
    
    def __init__(self, **kwargs):
        super(Swish, self).__init__(**kwargs)
        self.__name__ = 'Swish'
    def __call__(self, inputs):
        beta = 1.25
        return beta * inputs * sigmoid(inputs)
    def get_config(self):
        return super().get_config()

set_learning_phase(0)
model = load_model(r'path/to/model.h5', custom_objects={'eswish':eswish})
#run inference on model
model.predict()

Here is the specification for the 'old' environment:
Python 3.7
Tensorflow 2.6
Keras 2.6

Now, I want to run inference with this model in 'new' environment. Specifcation for the 'new' environment.
Python 3.12
Tensorflow 2.16.1
Keras 3.0

I understand that loading old model with keras.model.load_model is no longer usable in Keras 3. So currently I am trying to export the model's config and weights to model.json and model.weights.h5 seperately then load these files in new Keras 3.0 environment.
The problem is that the model has custom activation function and layers. I have the definition for these custom activation function and layers. The question is how can I export the config of the model to .json with these custom functions and layers? Here's the code snippet for what I'm trying to do.

In 'old' Keras 2. environment:

import json
from tensorflow.keras.backend import set_learning_phase
from tensorflow.keras.models import load_model

class eswish(Activation):
    """
    Custom Swish activation function for Keras.
    """
    ...

set_learning_phase(0)
model = load_model(r'path/to/model.h5', custom_objects={'eswish':eswish})

#export config to json with custom activation function?
json_config = model.to_json()

with open('model_config.json', 'w') as outfile:
    json.dump(json_config, outfile)

[dhantule]

Hi @demybug123,
Thanks for reporting this, similar issue #20081 . Could you try loading your model with tensorflow 2.15, save it with .keras and then try loading it with keras 3 and pass custom_objects arguments directly while loading.

[demybug123]

Hi @dhantule,
Thanks for your response. For my project, I ended up just downgrading the environment to Keras 2.6 and Python 3.9 instead of Keras 3.0 and Python 3.12 before I saw your message. But I did try your suggestion and it worked well. Thanks for the suggestion, I'll keep it in mind the next time I use Keras.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No