This repository has been archived by the owner on Feb 26, 2024. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 0
/
QDTRACK_PCAN_SAILVOS.txt
128 lines (104 loc) · 7.15 KB
/
QDTRACK_PCAN_SAILVOS.txt
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
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
copy the entire folder /beegfs/work/shared/kangdong_shared/env/qdtrack_joint/
to your own conda environment root, e.g. /home/yourname/.conda/envs/
use following commands to set up the environment:
module load anaconda
source activate qdtrack_joint
cd /beegfs/work/yourname/amodal-qdtrack (your local root of this repo)
python setup.py develop
cd mmtracking-0.5.1
python setup.py develop
1. (Amodal)QDTrack-mots
(1) train (Amodal)QDTrack-mots on SAIL-VOS(cut) from scatch:
in ./tools/train.py:
1. set "--config" as "../configs/QDTrack_mots.py"
2. set the "--work-dir" as the dir you wish
3. uncomment line 31 for training without validation (it's better to use train with val after the first two epoches, before that OOM could somehow happen)
in ./configs/QDTrack_mots.py:
1. using Resnet-50: comment lines 7~12,226,255 and uncomment lines 225,254 (default is ResNeXt101)
2. adjust lines 87 & 88 for different backbones
3. choose "data_root" at lines 42 & 43 as the path for SAIL-VOS or SAIL-VOScut
4. set ann_file & img_prefix under data.train & data.val for amodal/modal SAIL-VOS/SAIL-VOScut train & val
cd ./tools , sbatch slurm_train.sh
(2) test (Amodal)QDTrack-mots on the SAIL-VOS(cut):
in ./tools/test.py:
1. set "--config" as "../configs/QDTrack_mots.py"
2. set the "--checkpoint" as the .pth file you want to test
3. set the "--out" as the place to save the output .pkl file (switch lines 22 & 23 if not needed)
4. uncomment the line 47 to set a place for visualization if needed (test speed will reduce a lot)
5. set a threshold value at line 59 for '--show_score_thr'
in ./configs/QDTrack_mots.py:
1. choose "data_root" at lines 42 & 43 as the path for SAIL-VOS or SAIL-VOScut
2. set ann_file & img_prefix under data.test for amodal/modal SAIL-VOS/SAIL-VOScut test
cd ./tools , sbatch slurm_test.sh
2. (Amodal)PCAN
(1) train (Amodal)PCAN on SAIL-VOS(cut):
in ./tools/train.py:
1. set "--config" as "../configs/QDTrack_PCAN_fixed_refine.py"
2. set the "--work-dir" as the dir you wish
3. uncomment line 31 for training without validation (it's better to use train with val after the first two epoches, before that OOM could somehow happen)
in ./configs/QDTrack_PCAN_fixed_refine.py:
1. set the (Amodal)QDTrack-mots weight file at line 190 (here (Amodal)QDTrack-mots works as detector for PCAN)
2. if the used weight is trained from Resnet50, comment lines 7~12 in ../configs/QDTrack_mots.py
3. choose "data_root" at lines 37 & 38 as the path for SAIL-VOS or SAIL-VOScut
4. set ann_file & img_prefix under data.train & data.val for amodal/modal SAIL-VOS/SAIL-VOScut train & val
cd ./tools , sbatch slurm_train.sh
(2) test (Amodal)PCAN on the SAIL-VOS(cut):
in ./tools/test.py:
1. set "--config" as "../configs/QDTrack_PCAN_fixed_refine.py"
2. set the "--checkpoint" as the .pth file you want to test
3. set the "--out" as the place to save the output .pkl file (switch lines 22 & 23 if not needed)
4. uncomment the line 47 to set a place for visualization if needed (test speed will reduce a lot)
5. set a threshold value at line 59 for '--show_score_thr'
in ./configs/QDTrack_PCAN_fixed_refine.py:
1. choose "data_root" at lines 37 & 38 as the path for SAIL-VOS or SAIL-VOScut
2. set ann_file & img_prefix under data.test for amodal/modal SAIL-VOS/SAIL-VOScut test
cd ./tools , sbatch slurm_test.sh
3. QDTrack-mots-joint
(1) train QDTrack-mots-joint on SAIL-VOS(cut) from scatch:
in ./tools/train.py:
1. set "--config" as "../configs/QDTrack_mots_joint.py"
2. set the "--work-dir" as the dir you wish
3. uncomment line 31 for training without validation (it's better to use train with val after the first two epoches, before that OOM could somehow happen)
in ./configs/QDTrack_mots_joint.py:
1. set visible_mask_head loss & amodal_mask_head loss at lines 30 & 43
2. choose "data_root" at lines 55 & 56 as the path for SAIL-VOS or SAIL-VOScut
3. set ann_file & img_prefix under data.train & data.val for train & val on SAIL-VOS or SAIL-VOScut
cd ./tools , sbatch slurm_train.sh
(2) test QDTrack-mots-joint on the SAIL-VOS(cut):
in ./tools/test.py:
1. set "--config" as "../configs/QDTrack_mots_joint.py"
2. set the "--checkpoint" as the .pth file you want to test
3. set the "--out" as the place to save the output .pkl file (switch lines 22 & 23 if not needed)
4. uncomment the line 47 to set a place for visualization if needed (test speed will reduce a lot)
5. set a threshold value at line 59 for '--show_score_thr'
in ./configs/QDTrack_mots_joint.py:
1. choose "data_root" at lines 55 & 56 as the path for SAIL-VOS or SAIL-VOScut
2. set ann_file & img_prefix under data.test for test on SAIL-VOS or SAIL-VOScut
cd ./tools , sbatch slurm_test.sh
4. QDTrack-mots-joint+
(1) train QDTrack-mots-joint+ on SAIL-VOS(cut) from scatch:
in ./tools/train.py:
1. set "--config" as "../configs/QDTrack_mots_joint+.py"
2. set the "--work-dir" as the dir you wish
3. uncomment line 31 for training without validation (it's better to use train with val after the first two epoches, before that OOM could somehow happen)
in ./configs/QDTrack_mots_joint+.py:
1. set losses of visible_bbox_head & amodal_bbox_head & visible_mask_head & amodal_mask_head at lines 42~45 & 59 & 72
2. choose "data_root" at lines 84 & 85 as the path for SAIL-VOS or SAIL-VOScut
3. set ann_file & img_prefix under data.train & data.val for train & val on SAIL-VOS or SAIL-VOScut
cd ./tools , sbatch slurm_train.sh
(2) test QDTrack-mots-joint+ on the SAIL-VOS(cut):
in ./tools/test.py:
1. set "--config" as "../configs/QDTrack_mots_joint+.py"
2. set the "--checkpoint" as the .pth file you want to test
3. set the "--out" as the place to save the output .pkl file (switch lines 22 & 23 if not needed)
4. uncomment the line 47 to set a place for visualization if needed (test speed will reduce a lot)
5. set a threshold value at line 59 for '--show_score_thr'
in ./configs/QDTrack_mots_joint+.py:
1. choose "data_root" at lines 84 & 85 as the path for SAIL-VOS or SAIL-VOScut
2. set ann_file & img_prefix under data.test for test on SAIL-VOS or SAIL-VOScut
cd ./tools , sbatch slurm_test.sh
None:
when facing "RuntimeError: operation does not have an identity.", don't worry, just resume the previous epoch's
checkpoint at line 26 in ./tools/train.py to resume the training. (the cause for that is probably the limit of
the batch size, which is set to 4 when using the ResNeXt101 backbone, a bigger batch size is expected to eliminate
this problem of unstability in the training, however, the single A100 GPU's max memory of 40GB constaint that)