Error on demo.py.

[aliaksandrsiarohin]

Error on demo.py.

[kashi211]

command i used is:
python3 demo.py --config config/fashion-256.yaml --driving_video '11.mp4' --source_image '22.png' --checkpoint 'fashion.pth.tar' --relative --adapt_scale

[AliaksandrSiarohin]

Most likely you have wrong version of pytorch. You need torch==1.0.0

[yj7082126]

Hi. @AliaksandrSiarohin would you mind explaining this "bilinear warping" issue that occurs after pytorch 1.0.0 in detail? And in which part of the code does this issue occurs?

[AliaksandrSiarohin]

Hi,
The problem occurs only in training, at some point model start to diverge and produce some bad solutions. I've tried it with pytroch 1.1, maybe in later versions 1.3, 1.4 it does not occur.
Through I don't have time to investigate it deeply, I belive that the problem is in bilinear warping (

first-order-model/modules/generator.py

Line 57 in 720bdd8

return F.grid_sample(inp, deformation)

).
Also, at a test time there are no problems.

[yj7082126]

Thank you for your answer, @AliaksandrSiarohin . I tested this issue with versions 1.2, 1.3 and 1.4, and they all ended in vanishing gradients, but version 1.0.1 works. So I believe that some code change between 1.0.1 and 1.1 was the problem, yet on the update notes there wasn't a significant change on torch.nn.functional.

[AliaksandrSiarohin]

Dear @yj7082126 thank you for letting me know,
The reason, I believe it is in grid_sample because in my other project I meet the same problem when I add a grid_sample with PyTorch 1.1. Without it everything was fine.
I'm not familiar with a PyTorch code base, so it is hard to for me to understand what really changed.
Maybe when I have more time, I will try to remove this grid_sample and try different modes of it.
Anyway thank you for your investigation and for letting me know.

[yj7082126]

Thanks!