MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
iperovFaceswap and multi gpu training - fake
#1
some people like 'thefakening' think that faceswap trains the model in multi gpu mode.
But this is wrong.

Code:
model = multi_gpu_model(model, self.gpus)
it does not work actually, and heat your second card in vain. This is why I removed it from DFL.
my issue in keras team still without reply You are not allowed to view links. Register or Login to view.

To make multi gpu training requires a lot of complicated tensorflow-only code: gradient averager and per-gpu loss functions. Therefore multigpu training cannot be done by just adding model = multi_gpu_model(model, self.gpus)
#2
Cool of you to show up randomly after a long absence and shit on another project. very mature.
#3
(08-03-2019, 09:23 AM)iperov Wrote: You are not allowed to view links. Register or Login to view.some people like 'thefakening' think that faceswap trains the model in multi gpu mode.
But this is wrong.

Code:
model = multi_gpu_model(model, self.gpus)
it does not work actually, and heat your second card in vain. This is why I removed it from DFL.
my issue in keras team still without reply You are not allowed to view links. Register or Login to view.

To make multi gpu training requires a lot of complicated tensorflow-only code: gradient averager and per-gpu loss functions. Therefore multigpu training cannot be done by just adding model = multi_gpu_model(model, self.gpus)

good to see you around bro.  Hope you stick around and keep us plugged in with the latest and greatest.  Anyone really deepfaking knows that dfl shits on faceswap
#4
msymp:
iperov , can you please provide the full code and system environment you are currently using? Thanks.

iperov:
msymp ?? srsly ? why you need it.

The reason the "bug isnt fixed" is because iperov wont give them the info they need to sort it out lmao. what a dick.
#5
(08-03-2019, 09:23 AM)iperov Wrote: You are not allowed to view links. Register or Login to view.some people like 'thefakening' think that faceswap trains the model in multi gpu mode.
But this is wrong.

Code:
model = multi_gpu_model(model, self.gpus)
it does not work actually, and heat your second card in vain. This is why I removed it from DFL.
my issue in keras team still without reply You are not allowed to view links. Register or Login to view.

To make multi gpu training requires a lot of complicated tensorflow-only code: gradient averager and per-gpu loss functions. Therefore multigpu training cannot be done by just adding model = multi_gpu_model(model, self.gpus)

Iperov, you are wrong about multi-GPU in Faceswap.  Multi-gpu does not work in DeepFaceLab, removing it was your perogative as the developer of DFL.  Multi-GPU does work in Faceswap.  Different code base than DFL.  I know because I use multi-GPU in faceswap almost every day.  Multi-GPU in faceswap allows bigger batch sizes to be trained in one iteration and allows more complex models to be trained.
#6
(08-04-2019, 05:27 AM)JustAnnonymouse Wrote: You are not allowed to view links. Register or Login to view.
(08-03-2019, 09:23 AM)iperov Wrote: You are not allowed to view links. Register or Login to view.some people like 'thefakening' think that faceswap trains the model in multi gpu mode.
But this is wrong.

Code:
model = multi_gpu_model(model, self.gpus)
it does not work actually, and heat your second card in vain. This is why I removed it from DFL.
my issue in keras team still without reply You are not allowed to view links. Register or Login to view.

To make multi gpu training requires a lot of complicated tensorflow-only code: gradient averager and per-gpu loss functions. Therefore multigpu training cannot be done by just adding model = multi_gpu_model(model, self.gpus)

Iperov, you are wrong about multi-GPU in Faceswap.  Multi-gpu does not work in DeepFaceLab, removing it was your perogative as the developer of DFL.  Multi-GPU does work in Faceswap.  Different code base than DFL.  I know because I use multi-GPU in faceswap almost every day.  Multi-GPU in faceswap allows bigger batch sizes to be trained in one iteration and allows more complex models to be trained.
yeah hes just a dick.
#7
(08-04-2019, 05:27 AM)JustAnnonymouse Wrote: You are not allowed to view links. Register or Login to view.Multi-GPU in faceswap allows bigger batch sizes to be trained

no, it does not

also found a proof that keras does not support multi-gpu training, and gradients should be placed on each gpu and then averaged
You are not allowed to view links. Register or Login to view.
#8
(08-04-2019, 07:24 AM)iperov Wrote: You are not allowed to view links. Register or Login to view.
(08-04-2019, 05:27 AM)JustAnnonymouse Wrote: You are not allowed to view links. Register or Login to view.Multi-GPU in faceswap allows bigger batch sizes to be trained

no, it does not

also found a proof that keras does not support multi-gpu training, and gradients should be placed on each gpu and then averaged
You are not allowed to view links. Register or Login to view.

fuck off man. you literally came out of nowhere to shit on another project when you dont even have that feature implemented in your own code. Stay underground man, just cause you are the dev of a popular project doesnt mean you have to be an asshole to other projects.
#9
@dpfks why this man still not banned for breaking forum rules? i cannot understand why so many unreasonable insults from him
#10
(08-04-2019, 02:30 PM)iperov Wrote: You are not allowed to view links. Register or Login to view.@dpfks why this man still not banned for breaking forum rules? i cannot understand why so many unreasonable insults from him

iperov, why do you post complete nonsense about another project that you are not part of?  Feeling lonely these days?  It's very funny that you think insults should be a bannable offense.  You react to everyone who questions you in any way with insults and anger.  If unreasonable insults were a cause for a ban, you would be banned from every forum you ever posted in.

Forum Jump:

Users browsing this thread: 1 Guest(s)