Faceswap Forum

I suppose I earned "expert" the hard way. Torzdf is correct about the PCI lanes. The cards need to communicate with each other. The performance is hindered by the slowest cards. On typical mainboards the first PCIe slot will go 16x in the others are not populated. If the 'first' 2 are popu...

stepping-raz0r wrote: ↑Fri Dec 11, 2020 3:57 pm
Could you give an example...
(Based solely on faceswap considerations...)

Will say Nvidia whatever you do.
30x0 series isn't working quite yet. Nvidia drivers and some other software are not updated to support FS(among other things), but may in a couple months.

Yes , more time.
Also there are more variables to making a reasonable/high quality swap.

Number of training images on both sides.
Image quality of training data.
Which model.

Sure DFL is set up differently than FS. Especially things like tensor support, make comparing them not very easy.

With the condition of the same batch size, multi gpu verses single. Sometimes it will train only somewhat faster 120% or so.. What you can do, is increase your batch size to almost 2X of a single card. That's when you really see the benefits. Your EGs/sec will go way up. Maybe not 200% more like 160...

This is a curious question. I like it.

For the mole question, I would suspect it would do like beards and glasses, depending on frame and lighting, it would appear and disappear, maybe hang around faded.

For the entirety of your question.. who knows. Could be an interesting test.

Yes, restart

That looks just fine. Most mb split the pcie lanes to 8x8x You can find some that don't, I think, but those are the $500+ boards. It's also sometimes difficult to find boards with enough room to physically place 2-4 cards in simultaneously. On mine I have to remove one to remove the next, no clearan...

Of your 2 subjects, what is the the least amount of data, least photos? What model are you using? Go the biggest batch size you can, although I wouldn't go over a batch size of 80-100. Tends to actually make it train worse depending on model. Near the end of training (usually after 600K for me) , I ...

AMD doesn't focus on AI/compute. Highly unlikely they will make much changes to even begin to compete. Also there's an consideration that the software layer that allows AMD cards to do the compute will need to be updated. I did see a rumor that they were making some compute only cards instead of foc...

While I won't swear you're having the same issues I've had, one of my 2070's is factory overclocked a whopping (sarcasm) 50 MHz and gives me this error after a couple hours training. I just underclock them both to 1700mhz and it stops. Maybe it is a factory overclock, maybe not, butI agree with Torz...

There are many. I feel its faster due to the work on supporting the models with mixed precision .
The UI is way more intuitive.
The manual alignments tool is pretty sweet.
A bit more intuitive for sure.

Huh. Dunno.
If it was me, my next thoughts would be to be sure I'm using the right mask and alignments.
Check the convert settings if I happened to do something odd there.

If "A" was trained as the original actor and "b" the one you want to swap-in, it should.

Try activating swap model

: Screenshot from 2020-10-22 16-57-23.png (8.44 KiB) Viewed 15222 times

Short answer is when you don't personally see it, with your eyes , it getting any better. The graph is just measuring the loss so you have an idea of how quickly it's learning. But even when it looks near flat it's still learning and trying different combinations trying to get to "the best"...

Read thought the guides here viewforum.php?f=3&sid=c8b5c4286b5887100865db36488ba759

Come up with some 'test' and follow through following the instructions.

I've read through the guides 10 times and pick up stuff I missed.
95% of your questions are answered there.

Oh! thats almost my old computer. Had a 8300 and a 990fx mainboard.
Yea FS really didn't care for any multi gpu stuff.

I was able to do 2 separate training sessions just fine. Had 16Gb ram.

Other have reported windows doesn't report CUDA usage well.
I believe you can change what being monitored somewhere, but its not super important.
Edit: where it says 3D, copy, encode...ect....you can change those.
Other monitoring like gpuz or afterburner will show its working

With your setup, a 4 min startup vs my 2.3 min startup sounds reasonable. I have a x470 chipset. Doesn't rx/tx non stop at max rate every time NVtop displays sample, sure it jumps around. Although, most looked like this. Screenshot from 2020-10-18 05-20-33.png I feel the need to take care of what I'...

Ok , this was for some reason a painful test. Doesn't help with the startup ideas so much but. Question: Are you increasing your batch size when using distributed? Should allow roughly a 85% higher batch and EGs/sec goes up.. I'm grasping at straws with this one. Test Results: Villain mode, 2X2070, ...

Faceswap Forum

Search found 177 matches

Re: Distributed Training does not start

Re: Hardware best practices

Re: Need more details

Re: MULTI GPU - double speed at same batc size or not?

Re: MULTI GPU - double speed at same batc size or not?

Re: Understanding how variations in data impact results

Re: Double Eyebrow still

Re: What Do you think of this MB

Re: Smaller batch size = Worse training?

Re: Hardware best practices

Re: Error during extract - CUDA_ERROR_ILLEGAL_ADDRESS: an illegal memory access was encountered

Re: Question about 30xx support

Re: Conversion seems to swap Original > Original and not Original >Swap?

Re: Conversion seems to swap Original > Original and not Original >Swap?

Re: Optimal Convergence

Re: Best software/where to start

Re: Distributed with Dual 2060 supers

Re: GPU usage is only 2%

Re: Distributed with Dual 2060 supers

Re: Distributed with Dual 2060 supers