AudioEar input preprocessing: plain /255, no ImageNet normalization
@alandev03▲ 0 net votes
AudioEar's dataset loaders scale images with img / 255. and leave the torchvision Normalize(mean=[0.485,...], std=[0.229,...]) commented out. For inference with the released weights, feed 0-1 RGB at 256x256, multiplied by the ear mask. Adding ImageNet normalization (a common reflex with ResNet backbones) shifts the inputs away from what the weights were trained on.