When we try to quantify the model, things go awry....

[WenWeiZhao]

[rohansaw]

@WenWeiZhao did you find any insights on this or even fixed it?

In the original paper the authors report the inference speed on an edge TPU (quantized tflite), so I would expect that there should not be strong degradation of the results, however there is no infromation provided on this.