-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlaunch.sh
executable file
·68 lines (49 loc) · 4.78 KB
/
launch.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
#################
# EDM on CIFAR-10
#################
## Model training
python -m jcm.main --config configs/cifar10_k_ve.py --workdir cifar10-kve-score-loss --mode train
## Model evaluation (loss computation, sampling, and likelihood computation)
python -m jcm.main --config configs/cifar10_k_ve.py --workdir cifar10-kve-score-loss --mode eval --eval_folder eval
## Model metrics (FID & IS)
python -m jcm.main --config configs/cifar10_k_ve.py --workdir cifar10-kve-score-loss --mode metrics --eval_folder eval
######################################
# Consistency distillation on CIFAR-10
######################################
## CD with the l1 metric
python -m jcm.main --config configs/cifar10_ve_cd.py --workdir /path/to/cd-l1 --mode train --config.optim.lr=0.0002 --config.training.loss_norm='l1'
## CD with the l2 metric
python -m jcm.main --config configs/cifar10_ve_cd.py --workdir /path/to/cd-l2 --mode train --config.optim.lr=0.0004 --config.training.loss_norm='l2'
## CD with the LPIPS metric
python -m jcm.main --config configs/cifar10_ve_cd.py --workdir /path/to/cd-lpips --mode train --config.optim.lr=0.0004 --config.training.loss_norm='lpips'
## Sampling for CD (LPIPS)
python -m jcm.main --config configs/cifar10_ve_cd.py --workdir /path/to/cd-lpips --mode eval --config.optim.lr=0.0004 --config.training.loss_norm='lpips' --eval_folder eval --config.eval.begin_ckpt=80
## Computing FID for CD (LPIPS)
python -m jcm.main --config configs/cifar10_ve_cd.py --workdir /path/to/cd-lpips --mode metrics --config.optim.lr=0.0004 --config.training.loss_norm='lpips' --eval_folder eval --config.eval.begin_ckpt=80
######################################################
# Continuous-time consistency distillation on CIFAR-10
######################################################
## Continuous-time CD with the l2 metric
python -m jcm.main --config configs/cifar10_ve_continuous.py --workdir /path/to/cifar10-continuous-cd-l2 --mode train --config.optim.lr=0.001 --config.model.dropout=0.3 --config.training.loss_norm='l2'
## Continuous-time CD with stopgrad and the l2 metric
python -m jcm.main --config configs/cifar10_ve_continuous.py --workdir /path/to/cifar10-continuous-cd-l2-stopgrad --mode train --config.optim.lr=0.000005 --config.training.stopgrad=True --config.training.loss_norm='l2'
## Continuous-time CD with stopgrad and the LPIPS metric
python -m jcm.main --config configs/cifar10_ve_continuous.py --workdir /path/to/cifar10-continuous-cd-lpips-stopgrad --mode train --config.optim.lr=0.000005 --config.training.stopgrad=True --config.training.loss_norm='lpips'
##################################
# Consistency training on CIFAR-10
##################################
## CT with fixed N and mu, and the LPIPS metric
python -m jcm.main --config configs/cifar10_ve_ct_ema.py --mode train --workdir /path/to/ct-n-120-mu-0.99 --config.optim.lr=0.0004 --config.training.loss_norm='lpips' --config.model.num_scales=120 --config.training.dsm_target --config.training.target_ema=0.99
## CT with adaptive schedules and the LPIPS metric
python -m jcm.main --config configs/cifar10_ve_ct_adaptive.py --workdir /path/to/ct --mode train --config.optim.lr=0.0004 --config.training.loss_norm='lpips' --config.training.dsm_target --config.training.start_scales=2 --config.training.end_scales=150 --config.training.start_ema=0.9
## Sampling for CT
python -m jcm.main --config configs/cifar10_ve_ct_adaptive.py --workdir /path/to/ct --mode eval --eval_folder eval --config.optim.lr=0.0004 --config.training.loss_norm='lpips' --config.training.dsm_target --config.training.start_scales=2 --config.training.end_scales=150 --config.training.start_ema=0.9 --config.eval.begin_ckpt=74 --config.eval.end_ckpt=74
## FID computation for CT
python -m jcm.main --config configs/cifar10_ve_ct_adaptive.py --workdir /path/to/ct --mode metrics --eval_folder eval --config.optim.lr=0.0004 --config.training.loss_norm='lpips' --config.training.dsm_target --config.training.start_scales=2 --config.training.end_scales=150 --config.training.start_ema=0.9 --config.eval.begin_ckpt=74 --config.eval.end_ckpt=74
##################################################
# Continuous-time consistency training on CIFAR-10
##################################################
## Continuous-time CT with the l2 metric
python -m jcm.main --config configs/cifar10_ve_continuous.py --workdir /path/to/continuous-ct-l2 --mode train --config.optim.lr=0.000005 --config.training.loss_norm='l2' --config.training.dsm_target=True --config.training.stopgrad=True --config.training.finetune=True
## Continuous-time CT with the LPIPS metric
python -m jcm.main --config configs/cifar10_ve_continuous.py --workdir /path/to/continuous-ct-lpips --mode train --config.optim.lr=0.000005 --config.training.loss_norm='lpips' --config.training.dsm_target=True --config.training.stopgrad=True --config.training.finetune=True