Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
Thread Rating:
  • 3 Vote(s) - 5 Average
  • 1
  • 2
  • 3
  • 4
  • 5
mondomongerMondomonger's Base H128 Model for Deepfacelab
#31
(03-27-2019, 01:19 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.The newest version was released Mar.26, if you upgrade to that one, I don't think this model will work.


I was curious since I have a lot of old h128 models, so I pulled down today(March26)'s build. I was able to train a few iterations and run a convert without an issue.
#32
(03-27-2019, 02:01 AM)Endalus Wrote: You are not allowed to view links. Register or Login to view.
(03-27-2019, 01:19 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.The newest version was released Mar.26, if you upgrade to that one, I don't think this model will work.


I was curious since I have a lot of old h128 models, so I pulled down today(March26)'s build. I was able to train a few iterations and run a convert without an issue.

OK.... so I got a new AMD RX 470 8gb gpu (opencl) - I recon THIS will do the job? ON deepfacelab of course...
#33
(02-27-2019, 05:51 AM)dpfks Wrote: You are not allowed to view links. Register or Login to view.
(02-27-2019, 04:36 AM)GastonX Wrote: You are not allowed to view links. Register or Login to view.This model, along with your one for FakeApp (before I had FaceLab working) saves quite a bit of time. My first experiment was pretty much the hardest one, changing an Asian face to a Caucasian one and this model saved me several hours, the immediate zero-training result is a recognizable result better than what FakeApp can pull off after many hours. And by recognizable I mean it looks 90% like the source, with some minor problems/lack of clarity that more training fixes. Thanks for posting it.

Though I have a bug to report in case any of you happen to know why it occurred; Somewhere between 410K and 440K epochs during training when it dropped from a loss of roughly 0.05 to 0.015 it just went blank and skyrocketed to a loss of 8, remaining permanently blank. My settings were aside from the default were a batch size of 24 and pixel loss enabled. The batch of 24 resulted in the rare "out of memory but not an error" report, which is bizarre as FaceLab seems to use between 9600 and 9700MB of video memory no matter what I do, but I have an 11GB card. Does the system reserve the remaining 1.5GB or is that due to something from the FaceLab settings? Edit: I didn't notice it says right in the window how much memory is available, that 1.5GB is reserved it seems.

Sadly no way to fully use GPU vRAM on windows. Windows reserved ~20% to use. Only way is to use linux, or purchase the beast Tesla cards to use TCC mode.

Or... use windows 7/8, they don't steal GPU vram
#34
help me with the SAE convert parameters, I do not understand all the parameters and the faces are blurred at the end. what parameters I have to put and what numbers. can you make a detailed guide to lso parameters?

Forum Jump:

Users browsing this thread: 1 Guest(s)