From 09ef6d30966943db964a61fd889e98b44a6d41c4 Mon Sep 17 00:00:00 2001 From: iperov Date: Thu, 18 Aug 2022 00:06:52 +0400 Subject: [PATCH] update training instruction --- doc/user_faq/user_faq.md | 30 +++++++++++++++--------------- 1 file changed, 15 insertions(+), 15 deletions(-) diff --git a/doc/user_faq/user_faq.md b/doc/user_faq/user_faq.md index de9301b0..08d281bb 100644 --- a/doc/user_faq/user_faq.md +++ b/doc/user_faq/user_faq.md @@ -59,11 +59,11 @@ Gather 5000+ samples of your face with various conditions using webcam which wil Here public storage https://disk.yandex.ru/d/7i5XTKIKVg5UUg with facesets and models. -> Using pretrained "RTT model 224.zip" from public storage (see above) +> Using pretrained "RTT model 224 V2.zip" from public storage (see above) Make a backup before every stage ! -1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned +1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned 2. place your celeb to workspace/data_src/aligned @@ -71,21 +71,21 @@ Make a backup before every stage ! 4. replace dst faceset with your faceset in workspace/data_dst/aligned -5. continue train +500.000 +5. continue train +500.000, (optional) deleting inter_AB.npy every 100.000 (save, delete, continue run) 6. random_warp:OFF, train +500.000 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. finalize model by disabling masked training for 100-200 (not thousand) iterations. +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:Y, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default. +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. Make a backup before every stage ! -1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned +1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned 2. place your celeb to workspace/data_src/aligned @@ -101,7 +101,7 @@ 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. finalize model by disabling masked training for 100-200 (not thousand) iterations. +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. @@ -119,31 +119,31 @@ Src faceset is celebrity. Must be diverse enough in yaw, light and shadow condit Do not mix different age. The best result is obtained when the face is filmed from a short period of time and does not change the makeup and structure. Src faceset should be xseg'ed and applied. You can apply Generic XSeg to src faceset. -> Using pretrained "RTT model 224.zip" from public storage (see above) +> Using pretrained "RTT model 224 V2.zip" from public storage (see above) Make a backup before every stage ! -1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned +1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned 2. place your celeb to workspace/data_src/aligned 3. place model folder to workspace/model -4. do not change settings, train +500.000 iters +4. do not change settings, train +500.000 iters, + (optional) deleting inter_AB.npy every 100.000 (save, delete, continue run) -5. random_warp OFF, train +500.000, periodically (every 100.000 iters) disable masked training for 5.000 iters and enable again +5. random_warp OFF, train +500.000, + (optional) periodically (every 100.000 iters) disable masked training for 5.000 iters and enable again 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. finalize model by disabling masked training for 100-200 (not thousand) iterations. +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:Y, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default. +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. Make a backup before every stage ! -1. place RTM WF Faceset from public storage (see above) to workspace/data_dst/aligned +1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned 2. place your celeb to workspace/data_src/aligned @@ -155,7 +155,7 @@ 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. finalize model by disabling masked training for 100-200 (not thousand) iterations. +7. (optional) finalize model by disabling masked training for 100-200 (not thousand) iterations. > reusing trained SAEHD RTM model