Derivative for cudnn_grid_sampler_backward is not implemented.

[ChaerinMin]

Hi,

It seems that learn2learn does not support torch.nn.functional.grid_sample()'s derivative. If my understanding was right, are there any ways to readily deal with this problem or do you know any alternatives for grid_sample()? I appreciate that you have released such a wonderful implementation of meta-learning and find this greatly helpful for me.

Thank you again. I am looking forward to listening from you.

[ChaerinMin]

I have found another way to cope with this issue. 罗堃铭's strategy was useful. https://github.com/coolbeam/UPFlow_pytorch.git

[zslzx]

I have found another way to cope with this issue. 罗堃铭's strategy was useful. https://github.com/coolbeam/UPFlow_pytorch.git

hi, I find that 罗's code also contains grid_sample(), how to avoid that? could you introduce your solution in detail? Thanks a lot.

[zslzx]

Solved. I have found a reimplementation of grid_sample() that works.

[Ayiing]

Solved. I have found a reimplementation of grid_sample() that works.

hi, may I ask how to avoid grid_sample? I also meet the same problems,