forked from Superbia-X/FaithTrip
-
Notifications
You must be signed in to change notification settings - Fork 0
/
run.bat
57 lines (31 loc) · 2.6 KB
/
run.bat
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
@echo off
REM Setting Python Interpreter Path(your python path)
set python_path=C:\anaconda3\envs\FaithTrip\python.exe
REM Run FaithTrip with Fast-Sampling
python .\FaithTrip\train_diffusion.py --dataset Osak --lr 0.01 --batch_size 4 --d_model 16 --step 32 --max_noise 0.02 --sampling 2
python .\FaithTrip\train_diffusion.py --dataset Glas --lr 0.01 --batch_size 4 --d_model 64 --step 32 --max_noise 0.02 --sampling 2
python .\FaithTrip\train_diffusion.py --dataset Edin --lr 0.01 --batch_size 16 --d_model 64 --step 32 --max_noise 0.06 --sampling 4
python .\FaithTrip\train_diffusion.py --dataset Toro --lr 0.01 --batch_size 8 --d_model 64 --step 32 --max_noise 0.06 --sampling 4
REM Run FaithTrip Standard w/o Fast-Sampling to predict x_0
REM The FaithTrip set default values T(step) in 32 and beta_t(max_noise) in 0.02, respectively.
REM If you want to change it, try to add --step and --max_noise
python .\FaithTrip-S\train_diffusion.py --dataset Osak --lr 0.01 --batch_size 4 --d_model 16
python .\FaithTrip-S\train_diffusion.py --dataset Glas --lr 0.01 --batch_size 4 --d_model 64
python .\FaithTrip-S\train_diffusion.py --dataset Edin --lr 0.01 --batch_size 16 --d_model 64
python .\FaithTrip-S\train_diffusion.py --dataset Toro --lr 0.01 --batch_size 8 --d_model 64
REM Run FaithTrip-Z to predict noise
python .\FaithTrip-Z\train_diffusion.py --dataset Osak --lr 0.005 --batch_size 4 --d_model 32
python .\FaithTrip-Z\train_diffusion.py --dataset Glas --lr 0.005 --batch_size 4 --d_model 32
python .\FaithTrip-Ztrain_diffusion.py --dataset Edin --lr 0.005 --batch_size 16 --d_model 32
python .\FaithTrip-Z\train_diffusion.py --dataset Toro --lr 0.005 --batch_size 8 --d_model 32
REM Run BASE and BASE-CM
REM PS: max f1(pairs-f1) represent the results of BASE-CM and total f1(pairs-f1) represent the results of BASE
python .\BASE\train_base.py --dataset Osak --lr 0.001 --batch_size 4 --d_model 128
python .\BASE\train_base.py --dataset Glas --lr 0.001 --batch_size 4 --d_model 128
python .\BASE\train_base.py --dataset Edin --lr 0.001 --batch_size 16 --d_model 128
python .\BASE\train_base.py --dataset Toro --lr 0.001 --batch_size 8 --d_model 64
REM Run BASE-WSE
python .\BASE-WSE\train_base.py --dataset Osak --lr 0.001 --batch_size 4 --d_model 128 --se_weight 5
python .\BASE-WSE\train_base.py --dataset Glas --lr 0.001 --batch_size 4 --d_model 128 --se_weight 5
python .\BASE-WSE\train_base.py --dataset Edin --lr 0.001 --batch_size 16 --d_model 128 --se_weight 5
python .\BASE-WSE\train_base.py --dataset Toro --lr 0.001 --batch_size 8 --d_model 64 --se_weight 5