suchiz

Accuracy is something easy to compute for basic classifiers, because it's only based on the True positive, False positive, True negative, and False negative. But when it comes to further...

It's always resized, so you don't need to worry about labelling. Concerning the pngs, I think it would work. I didn't try but in data/coco.py where the images are loaded....

As the warning says, il will interprete this as one single image of size 1024*1024 with 32 channels instead of 32 images of size 1024*1024 (1channel). Because I admit 1024*1024*32...