ffcv
ffcv copied to clipboard
Resize Interpolation
In torchvision's RandomResizeCrop and Resize allow the user to specify the interpolation. This isn't an option in ffcv's RandomResizedCrop nor is it documented which interpolation is being used. Could this be included / documented?
The default doesn't seem to be the same (torchvision's seems to be Bilinear whereas FFCV defaults to nearest) which can lead to artifacting in the images.
I spoke too soon, the mode is not NEAREST but INTER_AREA
from looking at this line which is probably a good choice for downsampling (see opencv docs but not for upsampling.
But yeah, I agree that the authors should put this as part of the docs