Unable to quantize the model to tflite

[alexiniestags]

Hello, First of all, thank you for the amazing work made in this repository. I have tried through multiple ways to export the model to TFLite, and the one that works to me was the following:

phi=0
_, prediction_model, = efficientdet(phi=phi, weighted_bifpn=True, num_classes=90)
prediction_model.load_weights('./weights/efficientdet-d{}.h5'.format(phi), by_name=True)
converter = tf2.lite.TFLiteConverter.from_keras_model(prediction_model)
converter.allow_custom_ops = True
converter.target_spec.supported_ops = [tf2.lite.OpsSet.TFLITE_BUILTINS_INT8, tf2.lite.OpsSet.TFLITE_BUILTINS,tf2.lite.OpsSet.SELECT_TF_OPS]
converter.optimizations = [tf2.lite.Optimize.DEFAULT]
converter.inference_type = tf2.quint8
converter.inference_input_type = tf2.quint8
tflite_model = converter.convert()

The thing is that, as you can see in the code, I have enabled the flags to quantize the model and so on, but the input type is still float32. I have also tried the tflite model on a mobile phone with the TFLite Object Detection App and it works, but it takes more than a second per image (SD845), and the input that it receives is float32, where with a MobileNetV2-SSD quantized takes around 50ms. Any idea of how to quantize the model? I have also tried to export it as a frozen graph and the result is the same. Note that the converted used is the tf2 one. I am using tensorflow 2.1.0 to export it.

Thanks in advance!

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[ankitchandracs]

Is there any update after this? Where did you put this code? How did you use the exported model for inference?