The thread here states that we can only quantise the backbone of object detection models. Is there by any chance of quantising the whole detection model as well? For example, by defining custom self.quant and self.dequant layers before the forward pass of the RPN and Box predictor modules ?
I managed to quantise the backbone, but it doesn’t give me much decrease in model size because the weights from the RPN and Box predictor modules are far greater than the backbone (mobilenetv2)