github/DeepFaceLive
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLive synced 2025-07-05 20:42:12 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

update training instruction

Browse source
This commit is contained in:
iperov 2022-08-29 12:53:53 +04:00
parent 3b95be43e0
commit 8c6ee38eb6
1 changed files with 3 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

14
doc/user_faq/user_faq.md
View file

@ -77,8 +77,6 @@ Make a backup before every stage !
7. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.
8. (optional) finalize model by disabling masked training for 100-200 (not thousand) iterations.
> Using SAEHD model from scratch.
res:224, WF, archi:liae-udt, ae_dims:512, e_dims:64, d_dims:64, d_mask_dims:32, eyes_mouth_prio:N, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default.
@ -101,9 +99,7 @@ Make a backup before every stage !
8. GAN 0.1 power, gan_dims:32, Train until the src loss value has not increased in the last 12 hours.
9. (optional) finalize model by disabling masked training for 100-200 (not thousand) iterations.
10. export the model in .dfm format for use in DeepFaceLive. You can also try ordering a deepfake model from someone in Discord or forum.
9. export the model in .dfm format for use in DeepFaceLive. You can also try ordering a deepfake model from someone in Discord or forum.
</td></tr>
<tr><td colspan=2 align="center">
@ -129,14 +125,12 @@ Make a backup before every stage !
3. place model folder to workspace/model
4. do not change settings, train +500.000 iters, + (optional) deleting inter_AB.npy every 100.000 (save, delete, continue run)
4. do not change settings, train +500.000 iters, + deleting inter_AB.npy every 100.000 (save, delete, continue run)
5. random_warp OFF, train +500.000, + (optional) periodically (every 100.000 iters) disable masked training for 5.000 iters and enable again
5. random_warp OFF, train +500.000
6. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.
7. (optional) finalize model by disabling masked training for 100-200 (not thousand) iterations.
> Using SAEHD model from scratch
res:224, WF, archi:liae-udt, ae_dims:512, e_dims:64, d_dims:64, d_mask_dims:32, eyes_mouth_prio:N, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default.
@ -155,8 +149,6 @@ Make a backup before every stage !
6. GAN 0.1 power, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.
7. (optional) finalize model by disabling masked training for 100-200 (not thousand) iterations.
> reusing trained SAEHD RTM model
Models that are trained before random_warp:OFF, can be reused. In this case you have to delete INTER_AB.NPY from the model folder and continue training from stage where random_warp:ON. Increase stage up to 2.000.000 and more iters. You can delete inter_AB.npy every 500.000 iters to increase src-likeness. Trained model before random_warp:OFF also can be reused for new celeb face.