Is it necessary to normalize pixel values if there are only black and white ones (nothing in between/no greys), before feeding into ResNet18 for classification?
IOW, is this transform necessary?
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
Note: I'm currently stacking each black and white image 3 times so it aligns with ResNet's RGB expectation.
the model is trained expecting values with 0 mean and some measured variance.
Thinking about your case, you would getting something like, eg, blue channel pixel with 1:
(1-0.485)/0.229 = 2.24
and for pixel with 0:
(0-0.485)/0.229 = -2.11
If you are using pre trained weights i would guess that yes, it is necessary, otherwise, you can measure it in your training accuracy. Anyway, if you are not sure, test the result with and without it.
