Good afternoon all, I have a MAC with i7 and RX 590, But a question arises, is 30-50 iterations at basic settings and Dlight the norm?
Intel macOS with AMD GPU - Training speed
Read the FAQs and search the forum before posting a new topic.
This forum is for discussing tips and understanding the process involved with Training a Faceswap model.
If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.
Please mark any answers that fixed your problems so others can find the solutions.
Re: Intel macOS with AMD GPU - Training speed
I assume you are talking about on an AMD GPU.... If so, then the answer is "I don't know". I'm fairly certain you could get this running on an Nvidia GPU with Mixed Precision enabled. This probably wouldn't work on AMD though, however I haven't tested. The best thing you could do to make sure is try running it at a super low batch size (1 or 2) and see if it starts training. If it does, you can then try to increase batch size from there.
As to your earlier question re: speeds. Apple-silicon and AMD will definitely run slower than Nvidia, just because Cuda is much more mature and optimized. There are benchmarks for Apple-Silicon vs Nvidia out there which confirm this.
My word is final
Re: Intel macOS with AMD GPU - Training speed
torzdf wrote: ↑Sat Sep 02, 2023 11:02 amI assume you are talking about on an AMD GPU.... If so, then the answer is "I don't know". I'm fairly certain you could get this running on an Nvidia GPU with Mixed Precision enabled. This probably wouldn't work on AMD though, however I haven't tested. The best thing you could do to make sure is try running it at a super low batch size (1 or 2) and see if it starts training. If it does, you can then try to increase batch size from there.
As to your earlier question re: speeds. Apple-silicon and AMD will definitely run slower than Nvidia, just because Cuda is much more mature and optimized. There are benchmarks for Apple-Silicon vs Nvidia out there which confirm this.
And in your experience. Is 6 GB for Dlight 256x the norm or is it very little for this type of model? Right now I'm talking about the NVidia 3060 in my laptop
Re: Intel macOS with AMD GPU - Training speed
I have not used DLight for a long time (I'm not a fan of unbalanced models to be honest, I think they underperform). As I said before, I think it will run with mixed precision enabled (and potentially a low batch size)
My word is final
Re: Intel macOS with AMD GPU - Training speed
And what model would you recommend with quality/performance balance? Just after reading some of your fags I came to the conclusion that the best in terms of quality is Dlight. But maybe there is a model that is not inferior directly worse than Dlight, but that at the same time requires less resources from the GPU?
Re: Intel macOS with AMD GPU - Training speed
I would try DFaker if starting out (very few configuration options, so easy to use).
If you are more experienced, then creating a custom model in Phaze-A would be the way to go.
My word is final