Choose one or several GPU idxs (separated by comma).
[CPU] : CPU
[0] : NVIDIA GeForce RTX 3090
[0] Which GPU indexes to choose? :
0
Press enter in 2 seconds to override model settings.
[4] Autobackup every N hour ( 0..24 ?:help ) : 8
8
[n] Write preview history ( y/n ?:help ) :
n
[2000000] Target iteration :
2000000
[n] Flip SRC faces randomly ( y/n ?:help ) :
n
[y] Flip DST faces randomly ( y/n ?:help ) :
y
[16] Batch_size ( ?:help ) :
16
[n] Masked training ( y/n ?:help ) : y
[y] Eyes and mouth priority ( y/n ?:help ) :
y
[y] Uniform yaw distribution of samples ( y/n ?:help ) :
y
[y] Blur out mask ( y/n ?:help ) :
y
[y] Place models and optimizer on GPU ( y/n ?:help ) :
y
[y] Use AdaBelief optimizer? ( y/n ?:help ) :
y
[y] Use learning rate dropout ( n/y/cpu ?:help ) :
y
[y] Enable random warp of samples ( y/n ?:help ) : n
[0.05] Random hue/saturation/light intensity ( 0.0 .. 0.3 ?:help ) :
0.05
[0.0] GAN power ( 0.0 .. 5.0 ?:help ) : ?
Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don't disable. The higher the value, the higher the chances of artifacts. Typical fine value is 0.1 [0.0] GAN power ( 0.0 .. 5.0 ?:help ) : 0.1 0.1 [32] GAN patch size ( 3-640 ?:help ) : 32 [16] GAN dimensions ( 4-512 ?:help ) : 16
[0.0] Face style power ( 0.0..100.0 ?:help ) : ?
Learn the color of the predicted face to be the same as dst inside mask. If you want to use this option with 'whole_face' you have to use XSeg trained mask. Warning: Enable it only after 10k iters, when predicted face is clear enough to start learn style. Start from 0.001 value and check history changes. Enabling this option increases the chance of model collapse.
[0.0] Face style power ( 0.0..100.0 ?:help ) :
0.0
[0.0] Background style power ( 0.0..100.0 ?:help ) : ?
Learn the area outside mask of the predicted face to be the same as dst. If you want to use this option with 'whole_face' you have to use XSeg trained mask. For whole_face you have to use XSeg trained mask. This can make face more like dst. Enabling this option increases the chance of model collapse. Typical value is 2.0
[0.0] Background style power ( 0.0..100.0 ?:help ) :
0.0
[lct] Color transfer for src faceset ( none/rct/lct/mkl/idt/sot ?:help ) : ?
Change color distribution of src samples close to dst samples. Try all modes to find the best.
[lct] Color transfer for src faceset ( none/rct/lct/mkl/idt/sot ?:help ) :
lct
[n] Enable gradient clipping ( y/n ?:help ) :
n
[n] Enable pretraining mode ( y/n ?:help ) :
n
Initializing models: 100%|###############################################################| 7/7 [00:03<00:00, 1.88it/s]
Loading samples: 100%|############################################################| 1158/1158 [00:02<00:00, 454.60it/s]
Sort by yaw: 100%|#################################################################| 128/128 [00:00<00:00, 4736.36it/s]
Loading samples: 100%|##############################################################| 971/971 [00:02<00:00, 483.61it/s]
Sort by yaw: 100%|#################################################################| 128/128 [00:00<00:00, 3119.29it/s]
================== Model Summary ===================
== ==
== Model name: RTM Liae-UD-256_SAEHD ==
== ==
== Current iteration: 513374 ==
== ==
==---------------- Model Options -----------------==
== ==
== resolution: 256 ==
== face_type: wf ==
== models_opt_on_gpu: True ==
== archi: liae-ud ==
== ae_dims: 256 ==
== e_dims: 64 ==
== d_dims: 64 ==
== d_mask_dims: 22 ==
== masked_training: True ==
== eyes_mouth_prio: True ==
== uniform_yaw: True ==
== adabelief: True ==
== lr_dropout: y ==
== random_warp: False ==
== true_face_power: 0.0 ==
== face_style_power: 0.0 ==
== bg_style_power: 0.0 ==
== ct_mode: lct ==
== clipgrad: False ==
== pretrain: False ==
== autobackup_hour: 8 ==
== write_preview_history: False ==
== target_iter: 2000000 ==
== random_src_flip: False ==
== random_dst_flip: True ==
== batch_size: 16 ==
== gan_power: 0.1 ==
== gan_patch_size: 32 ==
== gan_dims: 16 ==
== blur_out_mask: True ==
== random_hsv_power: 0.05 ==
== ==
==------------------ Running On ------------------==
== ==
== Device index: 0 ==
== Name: NVIDIA GeForce RTX 3090 ==
== VRAM: 21.17GB ==
== ==
====================================================
Starting. Target iteration: 2000000. Press "Enter" to stop training and save model.
[15:34:55][#516028][0534ms][0.4937][0.2940]
1.19:
昨天开启 GAN, 训练到今天早晨的效果
[0.05] Random hue/saturation/light intensity ( 0.0 .. 0.3 ?:help ) :
0.05
[0.1] GAN power ( 0.0 .. 5.0 ?:help ) :
0.1
[32] GAN patch size ( 3-640 ?:help ) :
32
[16] GAN dimensions ( 4-512 ?:help ) :
16
[0.0] Face style power ( 0.0..100.0 ?:help ) : ?
Learn the color of the predicted face to be the same as dst inside mask. If you want to use this option with 'whole_face' you have to use XSeg trained mask. Warning: Enable it only after 10k iters, when predicted face is clear enough to start learn style. Start from 0.001 value and check history changes. Enabling this option increases the chance of model collapse.
[0.0] Face style power ( 0.0..100.0 ?:help ) : 0.001
0.001
[0.0] Background style power ( 0.0..100.0 ?:help ) : ?
Learn the area outside mask of the predicted face to be the same as dst. If you want to use this option with 'whole_face' you have to use XSeg trained mask. For whole_face you have to use XSeg trained mask. This can make face more like dst. Enabling this option increases the chance of model collapse. Typical value is 2.0
[0.0] Background style power ( 0.0..100.0 ?:help ) : 2.0
2.0
训练了一会儿,每个迭代的时间变长了1倍 loss 下降一般般
Starting. Target iteration: 2000000. Press "Enter" to stop training and save model.
[10:59:57][#626909][0993ms][0.3595][0.1564]
[11:24:47][#628445][1020ms][0.3580][0.1565]
[11:49:47][#629974][0979ms][0.3577][0.1559]
[12:14:46][#631529][0936ms][0.3573][0.1560]
[12:39:47][#633090][1431ms][0.3563][0.1557]
[13:04:46][#634659][0981ms][0.3560][0.1556]
[13:29:47][#636231][0966ms][0.3557][0.1548]
[13:54:46][#637802][1005ms][0.3547][0.1544]
[14:19:46][#639370][0961ms][0.3544][0.1547]
[14:44:47][#640943][1148ms][0.3539][0.1543]
[15:09:47][#642517][1315ms][0.3541][0.1546]
[15:34:47][#644084][0972ms][0.3535][0.1544]
[15:49:43][#645024][0991ms][0.3580][0.1722]
要把GAN调大试试,现在的速度太慢了16:00
Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don't disable. The higher the value, the higher the chances of artifacts. Typical fine value is 0.1
值越高,出现伪影的机会就越高。
增加到 1 试试看
[23:32:40][#671308][1025ms][0.3468][0.1539]
[23:57:31][#672847][0991ms][0.3443][0.1538]
[00:22:30][#674404][1062ms][0.3504][0.1540]
[00:47:30][#675966][0979ms][0.3432][0.1517]
[01:12:30][#677528][0942ms][0.3443][0.1528]
[01:37:30][#679091][1013ms][0.3424][0.1529]
[02:02:30][#680652][1016ms][0.3443][0.1540]
[02:27:30][#682214][0993ms][0.3416][0.1511]
[02:52:30][#683776][0928ms][0.3426][0.1515]
[03:17:30][#685339][0976ms][0.3409][0.1523]
[03:42:31][#686902][1317ms][0.3377][0.1506]
[04:07:31][#688464][1032ms][0.3420][0.1502]
[04:32:30][#690026][0978ms][0.3443][0.1522]
[04:57:30][#691588][0928ms][0.3363][0.1501]
[05:22:31][#693150][0956ms][0.3350][0.1497]
[05:47:30][#694712][0967ms][0.3385][0.1497]
[06:12:30][#696274][0999ms][0.3408][0.1500]
[06:37:30][#697836][0954ms][0.3345][0.1497]
[07:02:30][#699399][0973ms][0.3369][0.1495]
[07:27:33][#700958][0949ms][0.3398][0.1502]
[07:52:30][#702505][1032ms][0.3385][0.1495]
[08:17:31][#704057][1427ms][0.3341][0.1486]
[08:42:30][#705613][0966ms][0.3395][0.1492]
[09:07:30][#707169][0973ms][0.3356][0.1482]
[09:32:30][#708722][0991ms][0.3413][0.1487]
[09:57:30][#710263][0960ms][0.3368][0.1485]
跑了一晚上 效果一般般 【20220120】 loss 下降缓慢,人脸更加模糊了,担心是不是训练坏了,把GAN调整为 0.2
[0.0] Face style power ( 0.0..100.0 ?:help ) : ?
Learn the color of the predicted face to be the same as dst inside mask. If you want to use this option with 'whole_face' you have to use XSeg trained mask. Warning: Enable it only after 10k iters, when predicted face is clear enough to start learn style. Start from 0.001 value and check history changes. Enabling this option increases the chance of model collapse.
[0.05] Random hue/saturation/light intensity ( 0.0 .. 0.3 ?:help ) :
0.05
[0.1] GAN power ( 0.0 .. 5.0 ?:help ) :
0.1
[32] GAN patch size ( 3-640 ?:help ) :
32
[16] GAN dimensions ( 4-512 ?:help ) :
16
[0.0] Face style power ( 0.0..100.0 ?:help ) : ?
Learn the color of the predicted face to be the same as dst inside mask. If you want to use this option with 'whole_face' you have to use XSeg trained mask. Warning: Enable it only after 10k iters, when predicted face is clear enough to start learn style. Start from 0.001 value and check history changes. Enabling this option increases the chance of model collapse.
[0.0] Face style power ( 0.0..100.0 ?:help ) : 0.001
0.001
[0.0] Background style power ( 0.0..100.0 ?:help ) : ?
Learn the area outside mask of the predicted face to be the same as dst. If you want to use this option with 'whole_face' you have to use XSeg trained mask. For whole_face you have to use XSeg trained mask. This can make face more like dst. Enabling this option increases the chance of model collapse. Typical value is 2.0
[0.0] Background style power ( 0.0..100.0 ?:help ) : 2.0
2.0
训练了一会儿,每个迭代的时间变长了1倍 loss 下降一般般
Starting. Target iteration: 2000000. Press "Enter" to stop training and save model.
[10:59:57][#626909][0993ms][0.3595][0.1564]
[11:24:47][#628445][1020ms][0.3580][0.1565]
[11:49:47][#629974][0979ms][0.3577][0.1559]
[12:14:46][#631529][0936ms][0.3573][0.1560]
[12:39:47][#633090][1431ms][0.3563][0.1557]
[13:04:46][#634659][0981ms][0.3560][0.1556]
[13:29:47][#636231][0966ms][0.3557][0.1548]
[13:54:46][#637802][1005ms][0.3547][0.1544]
[14:19:46][#639370][0961ms][0.3544][0.1547]
[14:44:47][#640943][1148ms][0.3539][0.1543]
[15:09:47][#642517][1315ms][0.3541][0.1546]
[15:34:47][#644084][0972ms][0.3535][0.1544]
[15:49:43][#645024][0991ms][0.3580][0.1722]