How can i downsample a tensor representing an image using Nearest/Bilinear interpolation?
I’ve tried using
torch.nn.Upsample with a size smaller than the original one, my outputs seem fine and i don’t get any errors. Are there any problems i’m not seeing with this kind of usage of Upsample?
Upsample uses F.interpolate as suggested.
We can check the source to see what’s actually doing: https://pytorch.org/docs/stable/_modules/torch/nn/functional.html#interpolate
The default mode ‘nearest’ doesn’t suit downscaling but ‘bilinear’ works fine (please correct me if I’m wrong!).
Otherwise, just use mode ‘area’ or adaptive_avg_pool2d directly. However, the latter seems a bit slower than ‘bilinear’.