https://github.com/huawei-noah/Pretrained-IPT

IPT is special in 2 ways:

1. It supports multiple scales. IPT models can have any number of scales, so a single model can be 2x, 3x, and 4x. I currently deal with this by simply picking the largest scale and using that.
2. IPT only support images 48x48. To overcome this limitation, the official arch wraps IPT in another class that "chops" any input image into pieces 48x48. Frankly, I do not fully understand this code, so I just copied it for now.