search

NVlabs / imaginaire

NVIDIA's Deep Imagination Team's PyTorch Library
Other
4.01k stars 450 forks source link

inference on fs-vid-to-vid #69

Closed SURABHI-GUPTA closed 2 years ago

SURABHI-GUPTA commented 3 years ago

@arunmallya @dumerrill @mjgarland hey.. thanks for sharing your code. I tried running your face-forensics pre-trained model on custom reference test data, however, the results are not good.

Can you suggest anything I should try?

https://user-images.githubusercontent.com/22388014/115153448-8e1f8880-a093-11eb-8731-82414e5795c6.mp4

DaddyWesker commented 3 years ago

Kinda the same problem here. WOuld glad to get any help too.

https://user-images.githubusercontent.com/10774222/137279545-270f5efe-944c-4238-93a0-db90ee321475.mp4

mingyuliutw commented 2 years ago

Source images are important for getting good motion transfer performance for few-shot vid2vid.

grewe commented 2 years ago

@SURABHI-GUPTA we are having problems with inference and was hoping you could share your config file and any tips --or if you saw this same error https://github.com/NVlabs/imaginaire/issues/112

grewe commented 2 years ago

@tcwang0509 any idea about this....using the new config file from issue 106 we got from you results in this new issue and we have been through all the code and using even same data from training and fails. Please advise

grewe commented 2 years ago

If step_count is a variable that pytorch shoudl be supporting ---there is at least some kind of conflict and we are using the same version as was used recently in discussions with arun, etc.

grewe commented 2 years ago

If instead _step_count is a variable that should be maintained in the imaginaire code, we can not find any place it is initialized from doing a backward search from the error into the code