(dkby) DFL learning notes

One 、DFL Change your mind ： First convert the video into pictures , Extract faces from pictures , Learn features from faces . Then apply the model , First change the face of the picture , And then combine the pictures into a video , At the same time, bring the audio track of the original video .

Two 、 technological process ：
1.src Video decomposition image  2.dst Video decomposition image
3.src Extract the face  4.dst Extract the face
5. Training models  6. Application model  7. Synthetic video

3、 ... and 、 Script ：
2) src Video extraction image extract images from video data_src.bat
3) dst Video extraction image （ Full frame rate ） extract images from video data_dst FULL FPS.bat
4) src Automatically extract face data_src faceset extract.bat
5) dst Automatically extract face data_dst faceset extract.bat
6) Training SAEHD train SAEHD.bat
7) application SAEHD merge SAEHD.bat
8) synthesis MP4 video merged to mp4.bat

Four 、 Pre training model ：
If you use someone else's model ：use pretrain mode Change it to N, Turn off the pre training mode so that the model will restore the iteration count to 0
If you train the model yourself ： Create the model normally , Pre training mode selection n, then src and dst Of aligned In the folder , Put your existing face data pictures ,src and dst It doesn't matter to repeat , The more pictures, the better , The more complicated the better .

5、 ... and 、 Training parameters
1.[0] Which GPU indexes to choose? : Please enter GPU Serial number
2.Choose one of saved models, or enter a name to create a new model. Select an existing model , Or enter a name to create a new model
3.Choose one or several GPU idxs (separated by comma). Choose one or more GPU Serial number
4.[0] Autobackup every N hour Every time N Hour backup ( 0..24 ?:help ) : Select the automatic backup time
5.[n] Write preview history ( y/n ?:help ) : Record save preview
6.[n] Choose image for the preview history ( y/n ) : Specify the preview
7.[0] Target iteration Target iterate : Target iterations
8.[y] Flip faces randomly ( y/n ?:help ) : Flip the face randomly
9.[2] Batch_size ( ?:help ) : ? Batch size
10.[512] Resolution ( 64-640 ?:help ) : ? Model resolution
11.[wf] Face type Face type ( h/mf/f/wf/head ?:help ) : Select face type
12.[liae-ud] AE architecture ( ?:help ) : ?
'df' keeps more identity-preserved face.
'liae' can fix overly different face shapes.
'-u' increased likeness of the face.
'-d' (experimental) doubling the resolution using the same computation cost.
Examples: df, liae, df-d, df-ud, liae-ud, ...
The architecture of the model
13.[256] AutoEncoder dimensions ( 32-1024 ?:help ) : The width of the bottleneck layer in the middle of the model
14.[64] Encoder dimensions ( 16-256 ?:help ) : The width of the model coding layer
15.[64] Encoder dimensions ( 16-256 ?:help ) : The width of the model decoding layer
16.[n] Masked training ( y/n ?:help ) : ? Only train the mask part
17.[0.0] GAN power ( 0.0 .. 1.0 ?:help ) :GAN The intensity of

6、 ... and 、 Training and debugging parameters （2022.7.18 edition ）
face type : WF
random_flip : off
adabelief : on
eyes_mouth_prio : on
ct_mode: lct .

10-30W
learning rate drop：y

10-30W
random warp：n

GAN Training 10-30W
GAN poewer： 0.1