MAT: Mask-Aware Transformer for Large Hole Image Inpainting (CVPR 2022 Best Paper Finalist, Oral)

This work is base on the MAT, but I remove the speical .cu .cpp files write for torch 1.7-1.9.

The origin version may have error when you try to run the model in some higher torch version. even in torch1.7, it also have some to compile the .cpp files.

some comperation

1. as I use the general operation, it may cost more memory of GPU, as I use the batchsize of 4 in 256x256, it cost 2-3 gig more compare to the origin.
2. It run slower, you may have to cost 25% more time to train the model.
3. the only advantage is you can use it at any version of pytorch(>1.7), and the result is all as same as the origin.