I am working with images with pixel values -1, 0 or 1. The learned weights of the filter however are float values that lie between -1 and +1.
Is there anyway in Pytorch I can constrain the kernel filters to have entries -1, 0 or 1 to observe better learning capabilities?