388964e8d0
fix model export. Update tf2onnx to 1.9.2
2021-11-20 16:05:44 +04:00
853312e84c
fix
2021-10-18 13:18:59 +04:00
71f22957a6
SAEHD: random scale increased to -0.15+0.15. Improved lr_dropout capability to reach lower value of the loss.
...
When random_warp is off, inter_AB network is no longer trained to keep the face more src-like.
added option Random hue/saturation/light intensity applied to the src face set only at the input of the neural network. Stabilizes color perturbations during face swapping. Reduces the quality of the color transfer by selecting the closest one in the src faceset. Thus the src faceset must be diverse enough. Typical fine value is 0.05.
AMP: random scale increased to -0.15+0.15. Improved lr_dropout capability to reach lower value of the loss.
2021-10-18 11:03:23 +04:00
f48e852de3
SAEHD: better training for bg_style_power
2021-10-12 15:17:41 +04:00
b1b5d6f482
AMP, SAEHD: In the sample generator, the random scaling was increased from -0.05+0.05 to -0.125+0.125 , which improves the generalization of faces.
2021-10-09 13:58:46 +04:00
9e0079c6a0
SAEHD, AMP: removed the implicit function of periodically retraining last 16 “high-loss” samples
2021-09-29 16:48:54 +04:00
8a897f236f
fix blur_out_mask for tf 1.13, fix export model for dx12 live build.
2021-09-11 00:00:15 +04:00
bd0b408a9b
fix
2021-09-06 16:11:23 +04:00
335da71db6
saehd: remove sample processor count limitation
2021-09-06 14:16:19 +04:00
70c9ee9463
change face_style_power
2021-09-06 14:11:08 +04:00
d301a4799f
_
2021-09-06 14:07:20 +04:00
b1990d421a
AMP: don't train inters with random_warp-off
2021-09-03 13:50:31 +04:00
3fe8ce86b1
AMP: enable inter_src+inter_dst training
2021-08-30 18:52:29 +04:00
f346e35e7f
fix
2021-08-30 12:57:19 +04:00
6e094d873d
SAEHD: removed 'dst_denoise' option. Added -t arhi option.
2021-08-29 11:48:55 +04:00
01f1a084b4
AMP: remove dst_denoise option
2021-08-29 11:46:51 +04:00
e53be5e22d
fix
2021-08-24 13:18:51 +04:00
91187ecb95
AMP: some arhi change results training stabilization. New options blur_out_mask + rtm_dst_denoise. Sample processors count are no more limited to 8, thus if you have AMD processor with 16+ cores, increase paging file size.
2021-08-24 12:58:23 +04:00
8e63666390
SAEHD: add 'Blur out mask' and 'Denoise DST faceset' options.
2021-08-20 17:07:44 +04:00
c5584fbda0
XSeg: changed pretrain mode
2021-08-12 17:01:38 +04:00
55b947eab5
XSeg: added pretrain option.
2021-07-30 17:24:21 +04:00
83b1412da7
Export AMP/SAEHD: added "Export quantized" option
2021-07-30 17:13:46 +04:00
d26d579836
AMP change help of ct mode. Default inter dims now 1024
2021-07-29 23:14:30 +04:00
9e1bc5a153
AMP: return lr_dropout option
2021-07-28 11:00:19 +04:00
4be135af60
AMP: last high loss samples behaviour - same as SAEHD
2021-07-28 08:58:32 +04:00
336e0ca944
fix
2021-07-17 22:04:20 +04:00
0748b8d043
AMP: removed eyes mouth prio option(default enabled), removed masked training option(default enabled).
2021-07-17 21:58:57 +04:00
959a3530f8
AMP: default morph factor is now 0.5. Preview shows with 1.0
2021-07-17 10:27:07 +04:00
da8f33ee85
removing fp16 for amp
2021-07-15 13:23:33 +04:00
c9e0ba1779
removing fp16 for sae
2021-07-15 13:22:32 +04:00
7b8991a3cc
fix help
2021-07-15 00:50:34 +04:00
35877dbfd7
AMP, SAEHD: added option use_fp16: Increases training/inference speed, reduces model size. Model may crash. Enable it after 1-5k iters.
...
AMP: removed lr_dropout, now it is enabled by default;
2021-07-14 23:03:54 +04:00
ee1bc83a14
AMP, SAEHD: reverted GAN to december version.
2021-07-07 21:45:43 +04:00
446b95942f
fix
2021-07-01 15:08:11 +04:00
0444a9a565
default morph factor is now 0.1
2021-07-01 14:29:11 +04:00
075422e0c5
amp: fixes, removed pretrain option
2021-07-01 14:25:54 +04:00
63c794b3d0
rename dump_dflive to export_dfm
2021-06-28 18:26:54 +04:00
9d6b6feb1f
_
2021-06-28 14:27:09 +04:00
90a74efd89
fix
2021-06-26 23:06:30 +04:00
3c6bbe22b9
AMP: added Inter dimensions param. The model is not changed. Should be equal or more than AutoEncoder dimensions. More dims are better, but require more VRAM. You can fine-tune model size to fit your GPU.
2021-06-26 10:55:33 +04:00
5783191849
AMP: code refactoring, fix preview history
...
added dumpdflive command
2021-06-26 10:44:41 +04:00
62f1d57871
AMP: fix a merge bug where morph_factor was always = 1.0
2021-06-14 17:58:01 +04:00
5fac5ee4f3
AMP model: added ‘morph_factor’ option. [0.1 .. 0.5]
...
The smaller the value, the more src-like facial expressions will appear.
The larger the value, the less space there is to train a large dst faceset in the neural network.
Typical fine value is 0.33
AMP model: added ‘pretrain’ mode as in SAEHD
Windows build: Default pretrain dataset is updated with applied Generic XSeg mask
2021-06-01 09:37:55 +04:00
34b62862e0
fix help msg
2021-06-01 09:36:53 +04:00
e52b53f87c
AMP fix
2021-05-30 09:24:23 +04:00
11a7993238
fix
2021-05-29 22:23:07 +04:00
6f86c68e65
add AMP model
2021-05-29 18:57:35 +04:00
fdb143ff47
added AMD/Intel cards support via DirectX12 ( DirectML backend )
2021-04-22 18:19:15 +04:00
93fe480eca
revert
2021-04-19 13:35:14 +04:00
af0b3904fc
XSeg trainer: additional dssim loss.
2021-04-15 21:43:47 +04:00
iperov