Low GPU utilization
johnrom1243 opened this issue · 9 comments
I managed to make it work, however it's really slow at around 2.40 s/it and my GPU is hovering around 20 percent utilization, Im running a 3080 12gb
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Perhaps question was about video inference? If yes, then it will not actually work much to use the GPU - perhaps the same problem that was in SimSwap - one frame, processing, the next frame - processing. A possible solution is to implement batch processing of frames, operating on a batch size to avoid OOM. This would greatly reduce video processing time and increase GPU utilization accordingly.
Perhaps question was about video inference?
Ah yes that explains it. Currently the face detector can only handle batch size 1. Which sucks I know. I have plans to adopt this to handle an arbitrary batch size and try to make a buffer of frames that the FaceDancer can draw from. Which I can see being problematic solve when dealing with several faces that needs to be blended into the frame.
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Perhaps question was about video inference? If yes, then it will not actually work much to use the GPU - perhaps the same problem that was in SimSwap - one frame, processing, the next frame - processing. A possible solution is to implement batch processing of frames, operating on a batch size to avoid OOM. This would greatly reduce video processing time and increase GPU utilization accordingly.
Hi netrunner, how can I contact you privately for info? thanks in advance
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Perhaps question was about video inference? If yes, then it will not actually work much to use the GPU - perhaps the same problem that was in SimSwap - one frame, processing, the next frame - processing. A possible solution is to implement batch processing of frames, operating on a batch size to avoid OOM. This would greatly reduce video processing time and increase GPU utilization accordingly.
Hi netrunner, how can I contact you privately for info? thanks in advance
Hi, you can contact in my profile
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Perhaps question was about video inference? If yes, then it will not actually work much to use the GPU - perhaps the same problem that was in SimSwap - one frame, processing, the next frame - processing. A possible solution is to implement batch processing of frames, operating on a batch size to avoid OOM. This would greatly reduce video processing time and increase GPU utilization accordingly.
Hi netrunner, how can I contact you privately for info? thanks in advance
Hi, you can contact in my profile
On git there is no chance to contact you with a private message
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Perhaps question was about video inference? If yes, then it will not actually work much to use the GPU - perhaps the same problem that was in SimSwap - one frame, processing, the next frame - processing. A possible solution is to implement batch processing of frames, operating on a batch size to avoid OOM. This would greatly reduce video processing time and increase GPU utilization accordingly.
Hi netrunner, how can I contact you privately for info? thanks in advance
Hi, you can contact in my profile
On git there is no chance to contact you with a private message
There is an email address on my page. In general, it is better to ask questions related to FaceDancer directly to the author of the code, since I am not developer of this project.
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Perhaps question was about video inference? If yes, then it will not actually work much to use the GPU - perhaps the same problem that was in SimSwap - one frame, processing, the next frame - processing. A possible solution is to implement batch processing of frames, operating on a batch size to avoid OOM. This would greatly reduce video processing time and increase GPU utilization accordingly.
Hi netrunner, how can I contact you privately for info? thanks in advance
Hi, you can contact in my profile
On git there is no chance to contact you with a private message
There is an email address on my page. In general, it is better to ask questions related to FaceDancer directly to the author of the code, since I am not developer of this project.
Thanks but the topic is not about this project,anyway I don't see any email address maybe your profile is private
Hm, what batch size do you use? Could be some improvements that needs to be done. But when I trained my models it utilized the whole GPU with 1.1 it/s on batch size 10. That is a single 3090.
Perhaps question was about video inference? If yes, then it will not actually work much to use the GPU - perhaps the same problem that was in SimSwap - one frame, processing, the next frame - processing. A possible solution is to implement batch processing of frames, operating on a batch size to avoid OOM. This would greatly reduce video processing time and increase GPU utilization accordingly.
Hi netrunner, how can I contact you privately for info? thanks in advance
Hi, you can contact in my profile
On git there is no chance to contact you with a private message
There is an email address on my page. In general, it is better to ask questions related to FaceDancer directly to the author of the code, since I am not developer of this project.
Nothing to do I couldn't find your email anywhere.please send me your email address (vespersland@gmail.com)