Calling to_8bit on non-float input tensors produces incorrect results, likely due to how integer multiplication / division is handled. Likely fix is casting all inputs to float.
Code to replicate
x = torch.randint(0, 10000, (3, 10, 10))
output = to_8bit(x) # only 255 or 0
Calling
to_8biton non-float input tensors produces incorrect results, likely due to how integer multiplication / division is handled. Likely fix is casting all inputs to float.Code to replicate