In PyTorch there is iterpolate
.
I dug around the Caffe2 stack but the only thing which I could find is ResizeNearest
.
Now before you suggest using another function called ResizeLike
, ResizeLike
's behaviour is completely different to interpolate (it kinda works like transpose
)
ResizeNearest
isn’t too bad, least it allows for some dynamic forms of tensor resizing which changes with respect to input shapes.
But there are 2 issues (requires some knowledge of ResizeNearest
behavior)
- You are stuck with a resizing which always
floor
s your output dimension - You are forced to design around a constant scaling ratio
I hope this thread would be useful to many people because I’ve been trying to find a good way to go about this for a good day now…