-
Notifications
You must be signed in to change notification settings - Fork 0
/
scripts_training.sh
77 lines (55 loc) · 6.8 KB
/
scripts_training.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
69
70
71
72
73
74
75
76
77
# Copyright (c) 2017-2023 Mitsubishi Electric Research Laboratories (MERL)
#
# SPDX-License-Identifier: AGPL-3.0-or-later
#===============================================================================
# LUVLi Training Scripts
#===============================================================================
cd abhinav_model_dir
mkdir run_61 run_10{7..9} run_503 run_507 run_1001 run_1005 run_5000 run_5004
cd ..
# === 300-W Split 1 ====
python train_face_gll.py --gpu_id 0 --exp_id run_108 --pp "relu" --laplacian --use_visibility | tee abhinav_model_dir/run_108/train.log
# === 300-W Split 2 ====
# Train on 300W-LP-2D heatmaps
python train_face_gll.py --gpu_id 0 --exp_id run_61 --pp "" --train_json dataset/300W_LP_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --use_heatmaps --lr 0.00025 --lr_policy 3 --nEpochs 41 --saved_wt_file "" | tee abhinav_model_dir/run_61/train.log
# Finetune with 300W-LP-2D heatmap weights
python train_face_gll.py --gpu_id 0 --exp_id run_109 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar | tee abhinav_model_dir/run_109/train.log
# Finetune with 300-W heatmap weights
python train_face_gll.py --gpu_id 0 --exp_id run_107 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility | tee abhinav_model_dir/run_107/train.log
# === AFLW-19 ====
# Train on AFLW-19 heatmaps
python train_face_gll.py --gpu_id 0 --exp_id run_503 --pp "" --train_json dataset/aflw_train.json --val_json dataset/aflw_test_all.json --bulat_aug --use_heatmaps --lr 0.00025 --class_num 19 --lr_policy 3 --nEpochs 100 --saved_wt_file "" | tee abhinav_model_dir/run_503/train.log
# Finetune with AFLW-19 heatmap weights
python train_face_gll.py --gpu_id 0 --exp_id run_507 --pp "relu" --train_json dataset/aflw_train.json --val_json dataset/aflw_test_all.json --bulat_aug --laplacian --use_visibility --class_num 19 --saved_wt_file abhinav_model_dir/run_503/lr-0.0000125-99.pth.tar | tee abhinav_model_dir/run_507/train.log
# === WFLW ====
# Train on WFLW heatmaps
python train_face_gll.py --gpu_id 0 --exp_id run_1001 --pp "" --train_json dataset/wflw_train.json --val_json dataset/wflw_test.json --bulat_aug --use_heatmaps --lr 0.00025 --class_num 98 --lr_policy 3 --nEpochs 100 --saved_wt_file "" | tee abhinav_model_dir/run_1001/train.log
# Finetune with WFLW heatmap weights
python train_face_gll.py --gpu_id 0 --exp_id run_1005 --pp "relu" --train_json dataset/wflw_train.json --val_json dataset/wflw_test.json --bulat_aug --laplacian --use_visibility --class_num 98 --saved_wt_file abhinav_model_dir/run_1001/lr-0.0000125-99.pth.tar | tee abhinav_model_dir/run_1005/train.log
# === MERL-RAV (AFLW_ours) ====
# Train on MERL-RAV (AFLW_ours) heatmaps
python train_face_gll.py --gpu_id 0 --exp_id run_5000 --pp "" --train_json dataset/aflw_ours_all_train.json --val_json dataset/aflw_ours_all_val.json --bulat_aug --use_heatmaps --lr 0.00025 --lr_policy 3 --nEpochs 100 --saved_wt_file "" | tee abhinav_model_dir/run_5000/train.log
# Finetune with MERL-RAV (AFLW_ours) heatmap weights
python train_face_gll.py --gpu_id 0 --exp_id run_5004 --pp "relu" --train_json dataset/aflw_ours_all_train.json --val_json dataset/aflw_ours_all_val.json --bulat_aug --use_visibility --laplacian --saved_wt_file abhinav_model_dir/run_5000/lr-0.0000125-99.pth.tar | tee abhinav_model_dir/run_5004/train.log
#===============================================================================
# Ablation Studies
#===============================================================================
mkdir abhinav_model_dir/run_7{0..9}
# last hourglass wt
python train_face_gll.py --gpu_id 0 --exp_id run_70 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar --hg_wt "0,0,0,0,0,0,0,1" | tee abhinav_model_dir/run_70/train.log
# LUVLi to MSE
python train_face_gll.py --gpu_id 0 --exp_id run_71 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar --wt_mse 1 --wt_gau 0 | tee abhinav_model_dir/run_71/train.log
# LUVLi to UGGLI
python train_face_gll.py --gpu_id 0 --exp_id run_72 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar | tee abhinav_model_dir/run_72/train.log
# LUVLi to Gaussian Likelihood + visibility
python train_face_gll.py --gpu_id 0 --exp_id run_73 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar | tee abhinav_model_dir/run_73/train.log
# No visibility
python train_face_gll.py --gpu_id 0 --exp_id run_74 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar | tee abhinav_model_dir/run_74/train.log
# Get mean from MLP
python train_face_gll.py --gpu_id 0 --exp_id run_75 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar --get_mean_from_mlp | tee abhinav_model_dir/run_75/train.log
# relu --> smax 1
python train_face_gll.py --gpu_id 0 --exp_id run_76 --smax --tau 1 --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar | tee abhinav_model_dir/run_76/train.log
# relu --> smax 0.02
python train_face_gll.py --gpu_id 0 --exp_id run_77 --smax --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar | tee abhinav_model_dir/run_77/train.log
# 8 hourglass --> 4 hourglass
python train_face_gll.py --gpu_id 0 --exp_id run_78 --pp "relu" --train_json dataset/all_300Wtest_train.json --val_json dataset/all_300Wtest_val.json --bulat_aug --laplacian --use_visibility --saved_wt_file abhinav_model_dir/run_61/lr-0.00005-40.pth.tar --layer_num 4 --hg_wt "1,1,1,1" | tee abhinav_model_dir/run_78/train.log