How to Export Gemma 2 from Keras to Tensorflow Saved Model?

[JeeDevUser]

Just downloaded Gemma2 from
https://www.kaggle.com/models/google/gemma-2/keras

-unpacked and got the following files:

/assets
config.json
metadata.json
model.weights.h5
tokenizer.json

-tried to load model (and convert to Saved Model format) by using following Python script:

import json
from tensorflow import keras

# 1. Path to the config and weights:
config_path = "d:/Install/TensorFlow/models/Gemma_2/config.json"
weights_path = "d:/Install/TensorFlow/models/Gemma_2/model.weights.h5"
saved_model_dir = "d:/Install/TensorFlow/models/Gemma_2/gemma2_saved_model"

# 2. load model configuration
with open(config_path, 'r') as f:
    model_config = json.load(f)

# 3. Reconstruct the model:
model = keras.models.model_from_json(json.dumps(model_config))

# 4. Load the weights:
model.load_weights(weights_path)

# 5. Finally, save model as  TensorFlow SavedModel format:
model.save(saved_model_dir, save_format="tf")

print("Model saved to the :", saved_model_dir)

-at the line

model = keras.models.model_from_json(json.dumps(model_config))

got following error:

ValueError: Unknown layer: GemmaBackbone. Please ensure this object is passed to the custom_objectsargument. See https://www.tensorflow.org/guide/keras/save_and_serialize#registering_the_custom_object for details.

Questions:

• Can someone confirm that I'm on the right track?
• Looked at
  https://www.tensorflow.org/guide/keras/save_and_serialize#registering_the_custom_object

but not sure what to do with GemmaBackbone...huh?

Any help?

[JeeDevUser]

OK, I succeeded with following Python script:

import json
from tensorflow import keras
from keras_hub.models import GemmaBackbone

# Paths to configuration and weights
config_path = "d:/Install/TensorFlow/models/Gemma_2/config.json"
weights_path = "d:/Install/TensorFlow/models/Gemma_2/model.weights.h5"
saved_model_dir = "d:/Install/TensorFlow/models/Gemma_2/gemma2_saved_model"

# Load model configuration
with open(config_path, 'r') as f:
    model_config = json.load(f)

# Reconstruct the model with custom layer
custom_objects = {
    "GemmaBackbone": GemmaBackbone,
}
model = keras.models.model_from_json(json.dumps(model_config), custom_objects=custom_objects)

# Load the weights
model.load_weights(weights_path)

# Save the model as TensorFlow SavedModel
#model.save(saved_model_dir)
model.export(saved_model_dir)
print("Model saved to:", saved_model_dir)

BUT, I got an extra large model:
/variables/variables.data-00000-of-00001 - **20.917.805.737** in size
saved_model.pb - 4.056.008

Tried to load model, It ate up all 16GB of my memory, and failed to load.
Why is the Saved Model model so big, and is there a way to make it smaller?