Hi, I noticed that the image standardization in datagenerator.py prefetch is different from the implementation in match.py. I believe the way it's done in match.py is the correct one. I don't think the image should be divided by 255 before subtracting the mean and dividing by the standard deviation. I'd imagine this would impact performance quite a bit if the image standardization is different in training compared to testing?
Hi, I noticed that the image standardization in datagenerator.py prefetch is different from the implementation in match.py. I believe the way it's done in match.py is the correct one. I don't think the image should be divided by 255 before subtracting the mean and dividing by the standard deviation. I'd imagine this would impact performance quite a bit if the image standardization is different in training compared to testing?