When using F.interpolate with scale_factor not power of 2 (equal for both dimensions), during compilation it states the node is off-loaded (on accelerator), create a single graph (doesn’t cut the graph in this operation. But during inference on board it seems that it runs on the CPU and the stats accordingly – shows activity on both the accelerator and the CPU.
Can you explain the discrepancy? Does it happen with other type of layers or operations?
Thanks a lot
Avi.
