MobileNetV3 Quantised to int8 weights.

Prerequisites

Please answer the following question for yourself before submitting an issue.

• [x ] I checked to make sure that this feature has not been requested already.

1. The entire URL of the file you are using

https://github.com/tensorflow/models/blob/8bbed0227174cb6730e91cf8e922262ed2139ed3/research/slim/nets/mobilenet/README.md

2. Describe the feature you request

The README provides a mobile net v3 model quantised using uint8 weights. uint8 is deprecated by tflite-micro in favour of int8 OPs. It would be great if that is provided.

3. Additional context

I was looking to quantise the existing mobilenet_v3-small model with int8 weights. I could do that. The accuracy, however, makes it totally unusable.

The script I used is as below:

def representative_data_gen():
    """Generate calibration data with better coverage."""
    num_samples = 1000  # More calibration samples

    if not os.path.exists('train_images') or len(os.listdir('train_images')) < num_samples:
        train_images = download_sample_images(num_samples)

    for img_file in sorted(os.listdir('train_images'))[:num_samples]:
        img_path = os.path.join('train_images', img_file)
        img = load_img(img_path, target_size=(224, 224))
        img_array = img_to_array(img)

        # Generate multiple versions of each image
        variants = []

        # Original image
        img_array = img_array.astype(np.float32)
        img_normalized = img_array / 127.5 - 1
        variants.append(img_normalized)

        for variant in variants:
            yield [np.expand_dims(variant, axis=0).astype(np.float32)]

    # Load model
    model = tf.keras.applications.MobileNetV3Small(
        input_shape=(224, 224, 3),
        include_top=True,
        weights='imagenet',
        include_preprocessing=False
    )

    # Export model in SavedModel format
    print("Exporting model to SavedModel format...")
    model.export('mobilenet_v3_small_saved_model')

    # Convert from SavedModel format
    print("Converting from SavedModel format...")
    converter = tf.lite.TFLiteConverter.from_saved_model('mobilenet_v3_small_saved_model')

    # Basic quantization settings
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
    converter.representative_dataset = representative_data_gen

    # So, that it is easier to test against non-quantised model
    converter.inference_input_type = tf.float32
    converter.inference_output_type = tf.float32

    print("Converting model to TFLite...")
    tflite_model_quant = converter.convert()

    output_path = 'mobilenet_v3_small_quantized.tflite'
    with open(output_path, 'wb') as f:
        f.write(tflite_model_quant)

4. Are you willing to contribute it? (Yes or No)

No. Not well versed with the intricacies.

Looking for help which will make the quantisation work for MobileNetV3 small.