Faceswap Forum

Ryzen1988

I came across this YouTube channel that has some real clear short animated videos that explain a lot about the convolutional filtering part.

Can really recommend it for people who want some more understanding of the filtering steps.

Youtube

Ryzen1988

My latest dataset creation proces looked a lot like the previous mentioned, with a slight twist. I first sorted everything in 5 Yaw sorted groups, then did each group in 3 pitch groups. So you basically have looking up, straight and down in 5 groups from left to right. Then sorted each on blur in th...

Ryzen1988

the trick is just ignore the ugly strangeness for long enough en then it gets pretty

Ryzen1988

very interesting, thanks for the info.

Its a shame that all good things in life come at the cost of Vram

Ryzen1988

It took me a while to understand Lpips in usage because strange and ugly stuff seems to happen always. I went on with experimenting because my normal ms-ssim seemed to be not very effective in training with clipV So i did Logcosch + Lpips vgg16 and i was amazed. The first 3000-5000 its you only see ...

Ryzen1988

the clipv L encoder is certainly very suitable for 448-512px output.
The farl B i would personally use for 256-384px.
The clipv L - 334px is probably good for 512 and above but the encoder itself is already very massive so vram is a big issue, so whatever comes behind it as model needs to be skinny.

Ryzen1988

: learning_rate_finder_2023-09-25_14.34.57.png (33.74 KiB) Viewed 9468 times

That's what you would call a really hyperactive geek that's really wants to learn

Ryzen1988

Oke, i think i have narrowed it down to a specific setup. It happens with Clip farl B and Clip L when making a model with Hybrid for upscale and upsample and LCNR int activated. When replacing things for example with resize and upscale dyn and LCNR int off it has no slowdown issues, all other things...

Ryzen1988

Having trained a couple of different clip setups up and to around 200000 its, it has become clear that the clipv_vit-l-14 models really stand head and shoulder above the other ones, it's crazy almost every setup using one of the 2 vit-l models gives crazy fast and crazy good results. Only just a sho...

Ryzen1988

I have narrowed it down a bit, currently i'm training with Clipv_vit-l-14 and this seems to be without any issue. So the problem only happens when choosing the clipv-farl-b-16-64 so far. Loss functions don't influence the issue of slowdown. I will keep trying to a/b test to see if there is anything ...

Ryzen1988

I was premature, with your suggested solution training speed decays more slowly but it does decay still. Look at the Log, its every 1000 its saved, and it starts at 5 minute interval and slowly increases still 09/18/2023 18:23:29 INFO [Saved model] - Average loss since last save: face_a: 0.08979, fa...

Ryzen1988

I feel silly that fixed it.
By brain just hang on the fact that it was only with clipv

Ryzen1988

So i was trying to do a good training run with ClipV (slight modified 448 preset). Normally i calculate how many iterations can run in 24h by looking at the first 10 minutes x6x24. Big disappointment when looking, so i started to investigate. It seems that really in the first 30 minutes there happen...

Ryzen1988

very cool to hear that, should you ever be in need of high quality/high res faces sets for development or stuff i be more that happy to contribute.
Other than that I'm looking forward to these improvements.

Ryzen1988

+---------------------------------------------------------------------------------------+ | NVIDIA-SMI 537.13 Driver Version: 537.13 CUDA Version: 12.2 | |-----------------------------------------+----------------------+----------------------+ | GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. E...

Ryzen1988

So when reading at the Farlb github where the Clipv-farl transformer models come from. They also seem to have a ''State-of-the-art face alignment and face parsing model'' Would this possible be a good next gen solution instead of fan and s3fd and Bisenet FP combined? I have noticed that in swaps of ...

Ryzen1988

Another interesting curve

Ryzen1988

It was a relatively light and small model, 256px in -> 384px out batch size 16 if i am not mistaken, worked really well, so now i am making a 300px in -> 512px out with a way fatter FC. I will post what the performance/utilization with that when running as well "output_size": 384, "sp...

Ryzen1988

So i have noticed with the optimizer that the first 20% and last 20% is never where the useful part is. It takes a while before the Best rate start ticking up. Would it be hard to make the LR steps wider until the curve starts dropping, having dense small steps in the center of the curve but as soon...

Ryzen1988

Oke, oke i will run some tools for ya guys. I can say the 4090 really outperforms the a6000 (basicly a 3090) with a royal 40-50% Unfortunately the mem size is not up to par with the a6000 Wed Aug 30 19:13:32 2023 +--------------------------------------------------------------------------------------...

Faceswap Forum

Search found 57 matches

cool animated videos about CNN

Re: My way of creating good training set

Re: aaah i finally understand Lpips

Re: AutoClip Feedback from Personal AI models Outside FS

aaah i finally understand Lpips

Re: Yah new sutff, a vision transformer CLipV

Re: LR optimizer first impressions (yay new stuff)

Re: Training slowdown ClipV

Re: Yah new sutff, a vision transformer CLipV

Re: Training slowdown ClipV

Re: Training slowdown ClipV

Re: Training slowdown ClipV

Training slowdown ClipV

Re: Yah new sutff, a vision transformer CLipV

Re: Intel big-little headache

Re: Yah new sutff, a vision transformer CLipV

Re: LR optimizer first impressions (yay new stuff)

Re: Intel big-little headache

Re: LR optimizer first impressions (yay new stuff)

Re: Intel big-little headache