-
Notifications
You must be signed in to change notification settings - Fork 117
/
gmflow_scale2_regrefine6_train.sh
executable file
·170 lines (157 loc) · 4.88 KB
/
gmflow_scale2_regrefine6_train.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
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
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
#!/usr/bin/env bash
# GMFlow with hierarchical matching refinement (1/8 + 1/4 features)
# with additional 6 local regression refinements
# number of gpus for training, please set according to your hardware
# can be trained on 8x 32G V100 or 8x 40GB A100 gpus
NUM_GPUS=8
# chairs, resume from scale2 model
CHECKPOINT_DIR=checkpoints_flow/chairs-gmflow-scale2-regrefine6 && \
mkdir -p ${CHECKPOINT_DIR} && \
python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} --master_port=9989 main_flow.py \
--launcher pytorch \
--checkpoint_dir ${CHECKPOINT_DIR} \
--resume pretrained/gmflow-scale2-chairs-020cc9be.pth \
--no_resume_optimizer \
--stage chairs \
--batch_size 16 \
--val_dataset chairs sintel kitti \
--lr 4e-4 \
--image_size 384 512 \
--padding_factor 32 \
--upsample_factor 4 \
--num_scales 2 \
--attn_splits_list 2 8 \
--corr_radius_list -1 4 \
--prop_radius_list -1 1 \
--reg_refine \
--num_reg_refine 6 \
--with_speed_metric \
--val_freq 10000 \
--save_ckpt_freq 10000 \
--num_steps 100000 \
2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
# things
CHECKPOINT_DIR=checkpoints_flow/things-gmflow-scale2-regrefine6 && \
mkdir -p ${CHECKPOINT_DIR} && \
python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} --master_port=9989 main_flow.py \
--launcher pytorch \
--checkpoint_dir ${CHECKPOINT_DIR} \
--resume checkpoints_flow/chairs-gmflow-scale2-regrefine6/step_100000.pth \
--stage things \
--batch_size 8 \
--val_dataset things sintel kitti \
--lr 2e-4 \
--image_size 384 768 \
--padding_factor 32 \
--upsample_factor 4 \
--num_scales 2 \
--attn_splits_list 2 8 \
--corr_radius_list -1 4 \
--prop_radius_list -1 1 \
--reg_refine \
--num_reg_refine 6 \
--with_speed_metric \
--val_freq 40000 \
--save_ckpt_freq 50000 \
--num_steps 800000 \
2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
# sintel, resume from things model
CHECKPOINT_DIR=checkpoints_flow/sintel-gmflow-scale2-regrefine6 && \
mkdir -p ${CHECKPOINT_DIR} && \
python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} --master_port=9989 main_flow.py \
--launcher pytorch \
--checkpoint_dir ${CHECKPOINT_DIR} \
--resume checkpoints_flow/things-gmflow-scale2-regrefine6/step_800000.pth \
--stage sintel \
--batch_size 8 \
--val_dataset sintel kitti \
--lr 2e-4 \
--image_size 320 896 \
--padding_factor 32 \
--upsample_factor 4 \
--num_scales 2 \
--attn_splits_list 2 8 \
--corr_radius_list -1 4 \
--prop_radius_list -1 1 \
--reg_refine \
--num_reg_refine 6 \
--with_speed_metric \
--val_freq 20000 \
--save_ckpt_freq 20000 \
--num_steps 200000 \
2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
# sintel finetune, resume from sintel model, this is our final model for sintel benchmark submission
CHECKPOINT_DIR=checkpoints_flow/sintel-gmflow-scale2-regrefine6-ft && \
mkdir -p ${CHECKPOINT_DIR} && \
python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} --master_port=9989 main_flow.py \
--launcher pytorch \
--checkpoint_dir ${CHECKPOINT_DIR} \
--resume checkpoints_flow/sintel-gmflow-scale2-regrefine6/step_200000.pth \
--stage sintel_ft \
--batch_size 8 \
--val_dataset sintel \
--lr 1e-4 \
--image_size 416 1024 \
--padding_factor 32 \
--upsample_factor 4 \
--num_scales 2 \
--attn_splits_list 2 8 \
--corr_radius_list -1 4 \
--prop_radius_list -1 1 \
--reg_refine \
--num_reg_refine 6 \
--with_speed_metric \
--val_freq 1000 \
--save_ckpt_freq 1000 \
--num_steps 5000 \
2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
# vkitti2, resume from things model
CHECKPOINT_DIR=checkpoints_flow/vkitti2-gmflow-scale2-regrefine6 && \
mkdir -p ${CHECKPOINT_DIR} && \
python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} --master_port=9989 main_flow.py \
--launcher pytorch \
--checkpoint_dir ${CHECKPOINT_DIR} \
--resume checkpoints_flow/things-gmflow-scale2-regrefine6/step_800000.pth \
--stage vkitti2 \
--batch_size 16 \
--val_dataset kitti \
--lr 2e-4 \
--image_size 320 832 \
--padding_factor 32 \
--upsample_factor 4 \
--num_scales 2 \
--attn_splits_list 2 8 \
--corr_radius_list -1 4 \
--prop_radius_list -1 1 \
--reg_refine \
--num_reg_refine 6 \
--with_speed_metric \
--val_freq 10000 \
--save_ckpt_freq 10000 \
--num_steps 40000 \
2>&1 | tee -a ${CHECKPOINT_DIR}/train.log
# kitti, resume from vkitti2 model, this is our final model for kitti benchmark submission
CHECKPOINT_DIR=checkpoints_flow/kitti-gmflow-scale2-regrefine6 && \
mkdir -p ${CHECKPOINT_DIR} && \
python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} --master_port=9989 main_flow.py \
--launcher pytorch \
--checkpoint_dir ${CHECKPOINT_DIR} \
--resume checkpoints_flow/vkitti2-gmflow-scale2-regrefine6/step_040000.pth \
--stage kitti_mix \
--batch_size 8 \
--val_dataset kitti \
--lr 2e-4 \
--image_size 352 1216 \
--padding_factor 32 \
--upsample_factor 4 \
--num_scales 2 \
--attn_splits_list 2 8 \
--corr_radius_list -1 4 \
--prop_radius_list -1 1 \
--reg_refine \
--num_reg_refine 6 \
--with_speed_metric \
--val_freq 5000 \
--save_ckpt_freq 10000 \
--num_steps 30000 \
2>&1 | tee -a ${CHECKPOINT_DIR}/train.log