Rather than F.interpolate with optional alignment of corners you could try ResizeRight and in places where convolution strides are above 1 you can add Antialiased CNNs's BlurPool after them and their activation functions to make it more shift-invariant, hopefully increasing quality and making things more consistent. It's thankfully painless to implement.
Rather than F.interpolate with optional alignment of corners you could try ResizeRight and in places where convolution strides are above 1 you can add Antialiased CNNs's BlurPool after them and their activation functions to make it more shift-invariant, hopefully increasing quality and making things more consistent. It's thankfully painless to implement.