convert_keras_model_to_numpy.py

import os
import pickle

import numpy as np
import keras

from models import PointNetClassifierNumpy, numpy_layers


def convert_layer(keras_layer):
    """Convert"""

    # Get the type of layer
    layer_type = keras_layer.__class__.__name__
    # Find the equivalent (same naming)
    numpy_layer = getattr(numpy_layers, layer_type)
    # Initialize with data
    return numpy_layer.from_keras_layer(keras_layer)


if __name__ == "__main__":

    SAVE_DIR = "./models/saved/"

    # Load Keras model (choose among Keras saved models)
    trained_model = "PointNetClassifierModelNet40Full.keras"
    keras_model = keras.saving.load_model(os.path.join(SAVE_DIR, trained_model))

    # Convert to Numpy model
    numpy_model_layers = {layer.name: convert_layer(layer) for layer in keras_model.layers}
    numpy_model = PointNetClassifierNumpy(**numpy_model_layers)

    # Save Numpy model
    model_name = trained_model.replace(".keras", "Numpy.pkl")
    with open(os.path.join(SAVE_DIR, model_name), "wb") as f:
        pickle.dump(numpy_model, f)