The (image) datasets are often not normalized (e.g. pixel RGB encoded on unsigned bytes).
Currently data normalization is by dividing the data by 256 if input is not float32 type. But this is leading to a data range [0, 1]. This is not the range expected by some models.
Goal is to provide more normalizations with a GUI selector:
None
Mean 0, var 0 on batch
Range [0,1]: Detect uint8 input & divide by 256
Range [-1, 1]: Detect uint8 input & (data/128 - 1)
The (image) datasets are often not normalized (e.g. pixel RGB encoded on unsigned bytes).
Currently data normalization is by dividing the data by 256 if input is not float32 type. But this is leading to a data range [0, 1]. This is not the range expected by some models.
Goal is to provide more normalizations with a GUI selector: