Post-Quantization of mobilenetv2-keras model slower than your given quantized model

Hey,
I am quantizing the mobilenetv2_1.4_224 keras model using the post-training quantization tool (full integer). However, the runtime numbers are different to the runtime when I use your provided quantized mobilenetv2_1.4_224. Analyzing this issue, I found out that you converted it with TOCO, while by default the tflite converter uses MLIR. Explicitly setting converter.experimental_new_converter = False (use of TOCO) is in newer (>=2.6) TF versions not possible, because TOCO has been removed. Using previous TF version (<2.6) raise the error TypeError: ('Keyword argument not understood:', 'keepdims').

The difference between the full integer post-quantization using the TF guide and your provided quantized model is pretty severe. Your model runs with 4ms, while using the TF guide and MLIR it needs 10ms on the same HW.

Question: Could you please give me any information to reproduce your quantization of the mobilenetv2 ?

My quantization code is basically:

model = tf.keras.applications.mobilenet_v2.MobileNetV2(
    input_shape=(224,224,3), alpha=1.4, include_top=True, weights='imagenet',
    input_tensor=None, pooling=None, classes=1000,
    classifier_activation='softmax',)

    converter = tf.lite.TFLiteConverter.from_keras_model(model)
    converter.experimental_new_converter = True
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    converter.representative_dataset = representative_data_gen
    converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
    converter.inference_input_type = tf.uint8
    converter.inference_output_type = tf.uint8
    tflite_model = converter.convert()
    tflite_filepath.write_bytes(tflite_model)
`