MrDeepFakes Forums

Some content may not be available to Guests. Consider registering an account to enjoy unrestricted access to guides, support and tools

  • We are looking for community members who are intested in helping out. See our HELP WANTED post.

SAE model conversion seems to show only the original dst face.

tommo020788

DF Vagrant
Hi everyone, I just recently tried using the SAE training rather than the H128 training. Let it train for about 8 hours, came back, gave SAE conversion a try and the merged images seem to just be exactly the same as the original dst file with no changes made...
I've been following the settings found in the tutorial in the link below, but I'm open to the possibility I didn't follow it properly somehow.

Is there a common cause of this happening?
Link to the tutorial I'm following for SAE training:

https://pub.dfblue.com/pub/2019-07-27-deepfacelab-tutorial
 
F

Fappenberg

Guest
tommo020788 said:
Hi everyone, I just recently tried using the SAE training rather than the H128 training. Let it train for about 8 hours, came back, gave SAE conversion a try and the merged images seem to just be exactly the same as the original dst file with no changes made...
I've been following the settings found in the tutorial in the link below, but I'm open to the possibility I didn't follow it properly somehow.

Is there a common cause of this happening?
Link to the tutorial I'm following for SAE training:

https://pub.dfblue.com/pub/2019-07-27-deepfacelab-tutorial

Check data_dst\aligned_debug if the face is aligned.  If not, delete it and run 5) data_dst extract faces MANUAL RE-EXTRACT DELETED RESULTS DEBUG and manually align it.

Also i think you should run 5) data_dst extract faces MT all GPU + manual fix instead of the one listed in the guide, i believe MT is better than S3FD at detecting faces but with more false positives.  The manual fix will allow you to manually align it if it can't find a face.
 

tommo020788

DF Vagrant
fapper93 said:
tommo020788 said:
Hi everyone, I just recently tried using the SAE training rather than the H128 training. Let it train for about 8 hours, came back, gave SAE conversion a try and the merged images seem to just be exactly the same as the original dst file with no changes made...
I've been following the settings found in the tutorial in the link below, but I'm open to the possibility I didn't follow it properly somehow.

Is there a common cause of this happening?
Link to the tutorial I'm following for SAE training:

https://pub.dfblue.com/pub/2019-07-27-deepfacelab-tutorial

Check data_dst\aligned_debug if the face is aligned.  If not, delete it and run 5) data_dst extract faces MANUAL RE-EXTRACT DELETED RESULTS DEBUG and manually align it.

Also i think you should run 5) data_dst extract faces MT all GPU + manual fix instead of the one listed in the guide, i believe MT is better than S3FD at detecting faces but with more false positives.  The manual fix will allow you to manually align it if it can't find a face.
Thanks a lot for the reply. I'll give that a try today.
Gotta get this SAE working, cos it has all the features I want >:D
 

tommo020788

DF Vagrant
EDIT: Figured it out. Didn't notice there was a vram mem error, and what I was looking at was a collapsed model.
Reduced a couple settings and now its working :)

I'm using 1080ti graphics card.
 

ColeRogan

DF Vagrant
tommo020788 said:
EDIT: Figured it out. Didn't notice there was a vram mem error, and what I was looking at was a collapsed model.
Reduced a couple settings and now its working :)

I'm using 1080ti graphics card.

Where did you see the error?  What settings did you reduce? I'm having a similar if not exact problem.  Thanks
 

tommo020788

DF Vagrant
ColeRogan said:
tommo020788 said:
EDIT: Figured it out. Didn't notice there was a vram mem error, and what I was looking at was a collapsed model.
Reduced a couple settings and now its working :)

I'm using 1080ti graphics card.

Where did you see the error?  What settings did you reduce? I'm having a similar if not exact problem.  Thanks

I saw the error after it loaded a certain amount in the command window.
When I pulled up a GPU hardware monitor, I could see it was trying to use up just over 11gb of memory when the max my gpu is capable is 11gb (so it was going slightly over its max technical capacity).
The error would happen right after the gpu's memory was maxed out (not like ur typical popup crash error, but rather an error reported in the cmd window).
So all I had to do is change certain settings to reduce gpu usage.

The relevant settings I used were:
Resolution:__________128
Face_type: __________full
Learn_mask:__________y
Archi:_________________df
AE_Dims______________256
E_Ch_Dims:___________The Default Option/hit enter (the default had a lower setting). If you already have, u could try lowering it more.
D_Ch_dims____________Same as e_ch_dims.
Multi-scale_decoder:__y
Ca_weights:___________y
Face_style_power:____10 (u could also set this to 0 if u still have issues, but don't if u can manage not to).
Bg_style_power:______10 (see style power note above)

The Two settings that reduced gpu mem usage the most were: E_Ch_Dims and D_Ch_dims. I believe the trade off of lowering those two settings is that it would process training slower, but I'm not certain.
P.s. ^These two settings mentioned just above are abbreviations of two setting names, You'll notice which settings fit those abbreviations as you go through all the settings in the SAE training window, I just can't remember the full name of those settings off the top of my head XD

Hope that helps in some way. :D
 
Top