Search found 23 matches

by swapration
Sun Aug 22, 2021 7:26 pm
Forum: Training Discussion
Topic: Around what learning rate and epsilon should it be set for 384 px Phaze-A models using mixed precision?
Replies: 3
Views: 8618

Re: Around what learning rate and epsilon should it be set for 384 px Phaze-A models using mixed precision?

Do you happen to remember what average loss you were getting at 100k and 200k iterations? I started a new model that's at 200k iterations now, at an epsilon of -4. But while it seems much more stable - the loss troughs and peaks are no longer crazy far from the average - it also seems to have slowed...
by swapration
Wed Aug 18, 2021 12:56 am
Forum: Training Discussion
Topic: Around what learning rate and epsilon should it be set for 384 px Phaze-A models using mixed precision?
Replies: 3
Views: 8618

Around what learning rate and epsilon should it be set for 384 px Phaze-A models using mixed precision?

Around what learning rate and epsilon should it be set for 384 px Phaze-A models using mixed precision? Tried twice, training 384 px models. Got over 200k iterations both times until non stop NaNs kicked in. Setting epsilon to -5 or -4 doesn't help so I guess it was a mistake to start with an epsilo...
by swapration
Thu Jul 01, 2021 4:59 pm
Forum: Training Discussion
Topic: Why are models locked to specific mask types rather than say... just centering (legacy, face, head)?
Replies: 2
Views: 9144

Re: Why are models locked to specific mask types rather than say... just centering (legacy, face, head)?

Was looking over the code and was thinking... If there was a mixed mask option that allows you to change the mask type (under config. settings) and a range of frames to run the mask tool over wouldn't it theoretically allow for mixed mask models? For example the user first runs Bisenet-Fp for all fr...
by swapration
Thu Jul 01, 2021 4:43 pm
Forum: Convert Discussion
Topic: An idea for improving swapping differing face sizes
Replies: 1
Views: 9453

An idea for improving swapping differing face sizes

Was reading this post - https://forum.faceswap.dev/viewtopic.php?f=27&t=1647 - of someone trying to replace a large face with a smaller face and it got me thinking, more believable swaps could probably be made if a 'heal' pass (take some of the surrounding background around a face and drawn inwa...
by swapration
Thu Jul 01, 2021 4:56 am
Forum: Extract Discussion
Topic: Looks like Bisenet-Fp +Ears needs a larger extract box since ears get cut off
Replies: 2
Views: 13127

Re: Looks like Bisenet-Fp +Ears needs a larger extract box since ears get cut off

I was able to fix it by making the ears option use 'head' centering - editing line 34 in bisenet_fp.py to: self._storage_centering = "head" if self.config["include_hair"] or self.config["include_ears"] else "face" Tested out the segment options, it's cool how ...
by swapration
Mon Jun 28, 2021 3:08 am
Forum: Training Discussion
Topic: Loading weights and freezing layers on Phaze-A with split fc
Replies: 5
Views: 11435

Re: Loading weights and freezing layers on Phaze-A with split fc

Thank you for that! All the Phaze-A settings make so much more sense now. Although now other questions have come to mind... so much to test out.

by swapration
Sun Jun 27, 2021 11:03 pm
Forum: Training Discussion
Topic: Loading weights and freezing layers on Phaze-A with split fc
Replies: 5
Views: 11435

Re: Loading weights and freezing layers on Phaze-A with split fc

Oh, one last question.
Why is it when you reuse the encoder there's no identity leakage (as stated in the '[Guide] Training best practices...')?
Say you load and freeze the encoder weights from a face a1 + b1 model to a face a2 + b1 model then shouldn't face a1 data leak over?

by swapration
Sat Jun 26, 2021 7:58 pm
Forum: Training Discussion
Topic: Why are models locked to specific mask types rather than say... just centering (legacy, face, head)?
Replies: 2
Views: 9144

Why are models locked to specific mask types rather than say... just centering (legacy, face, head)?

Why are models locked to specific mask types rather than say... just centering (legacy, face, head)? Which can then determine the size of extraction, etc. Shouldn't the masking types just be a tool for generating specific masks for specific situations (obscured, side faces, hair, no hair, etc.)? Sor...
by swapration
Wed Jun 23, 2021 4:02 pm
Forum: Extract Discussion
Topic: Request: Manual tool option to grow/expand mask by # pixels like Gimp/Ps
Replies: 1
Views: 10747

Request: Manual tool option to grow/expand mask by # pixels like Gimp/Ps

Not sure if this is possible but it would be nice to be able to grow/expand or shrink/contract a mask you're editing like with image editing tools.

by swapration
Wed Jun 23, 2021 4:14 am
Forum: Extract Discussion
Topic: Looks like Bisenet-Fp +Ears needs a larger extract box since ears get cut off
Replies: 2
Views: 13127

Looks like Bisenet-Fp +Ears needs a larger extract box since ears get cut off

Bisenet-FP with ears option selected is currently broken since the ears get cut off as the face turns towards a side. Looks like it needs a larger extract box like Bisenet-Fp Head.

by swapration
Sun Jun 20, 2021 6:00 am
Forum: Training Discussion
Topic: Loading weights and freezing layers on Phaze-A with split fc
Replies: 5
Views: 11435

Loading weights and freezing layers on Phaze-A with split fc

With training new Phaze-A models that reuse Face B and replaces Face A, since only Face B is kept the same, with split fc enabled, in addition to the encoder should Fc B also be selected while Fc A left unselected? Also, while keeping only Face B the same and using a shared G-Block, under weights sh...
by swapration
Wed Jun 09, 2021 9:24 pm
Forum: Training Discussion
Topic: mask accuracy in relation to higher face coverage e.g. 87% legacy
Replies: 3
Views: 7227

Re: mask accuracy in relation to higher face coverage e.g. 87% legacy

I'm guessing by 'convert the mask needs to be correct' do you mean for face A, B, or both? I'm guessing you mean face A needs to be accurate due to obstructions?

by swapration
Wed Jun 09, 2021 9:19 pm
Forum: Training Discussion
Topic: "Head" option in modelsettings
Replies: 3
Views: 7340

Re: "Head" option in modelsettings

Was wondering how 'full head' models turn out as well. Due to all the extra detail I'm guessing you need to use a really high resolution model like sae set for max detail, something like 256 input and auto dim set greater than 512? Also would 'head' centering be best if you want to do face + ears? P...
by swapration
Sat Jun 05, 2021 10:45 pm
Forum: Training Discussion
Topic: mask accuracy in relation to higher face coverage e.g. 87% legacy
Replies: 3
Views: 7227

mask accuracy in relation to higher face coverage e.g. 87% legacy

Is it bad for masks to cover more than the edges of a face (backgrounds, hair, .etc) for models with higher coverage - for capturing a more accurate silhouette e.g. the bridge of the nose, cheekbone protrusion, etc.? Will it just add to the training time or will it draw things outside of the face d...
by swapration
Thu Mar 11, 2021 5:36 am
Forum: Hardware
Topic: Can a 2060s or 2060 run dfl-sae at 256 resolution?
Replies: 5
Views: 13503

Re: Can a 2060s or 2060 run dfl-sae at 256 resolution?

Thanks that clears up a lot.
Sounds like it really is best to go for 8gb at the minimum. You guys think the 3060, with it's 12gb, will turn out to be the best 'bang for your buck', once 30 series is fully supported?

by swapration
Wed Mar 10, 2021 12:07 am
Forum: Hardware
Topic: Can a 2060s or 2060 run dfl-sae at 256 resolution?
Replies: 5
Views: 13503

Can a 2060s or 2060 run dfl-sae at 256 resolution?

With mixed precision, is the 8gb/6gb vram of the 2060s/2060 enough for dfl-sae at a resolution of 256? Or is it not enough memory?
If it is enough to run it, what kind of EGs/sec does it run it at?

by swapration
Sun Sep 27, 2020 2:39 am
Forum: Training Discussion
Topic: Doing a short final run with warp disabled
Replies: 2
Views: 2005

Doing a short final run with warp disabled

Once we're satisfied with the details after say... 300k iterations of h128, how long of a run should we do with warp disabled?

by swapration
Fri Sep 18, 2020 7:10 am
Forum: Hardware
Topic: Theoretical performance of 30 vs 20 series cards.
Replies: 2
Views: 2878

Re: Theoretical performance of 30 vs 20 series cards.

There's some tensorflow benchmarks out for the 3080. Which can be found here: https://www.pugetsystems.com/labs/hpc/RTX3080-TensorFlow-and-NAMD-Performance-on-Linux-Preliminary-1885/#ComparativeCharts Looks like I failed to read the note below the chart. The 3080 was only tested against an RTX Titan...
by swapration
Thu Sep 17, 2020 12:30 am
Forum: Hardware
Topic: Theoretical performance of 30 vs 20 series cards.
Replies: 2
Views: 2878

Theoretical performance of 30 vs 20 series cards.

The 3070 is supposed to have equal FP16 and FP32 performance at 20 TFLops each compared to 2080 ti which has a higher FP16 of 26 but a lower FP32 at 13. Seems like a 'jack of all trades' vs 'master of one' scenario. Which theoretically should have a higher impact on performance in faceswap - 20 and ...