For the figure, the each step corresponds to a processed batch(16).
15k dataset
Resdcn + FPNx2 (↓:32,16,8,4)
18 layer
with warm-up and cosine schedule
1
2
3
4
5
6
| python train.py --log_name resdcn18_all \
--data_dir ~/cyl/Data/PSR_final \
--arch resdcn_18 \
--lr 1e-5 \
--batch_size 36 \
--num_epochs 90 --num_workers 0 --log_interval 5
|
### 50 layer
#### with warm-up and cosine schedule
1
2
3
4
5
6
| python train.py --log_name resdcn50_all \
--data_dir ~/cyl/Data/PSR_final \
--arch resdcn_50 \
--lr 1e-5 \
--batch_size 24 \
--num_epochs 90 --num_workers 0 --log_interval 5
|
### 101 layer
#### Use PSR dataset before remove bg samples (12415 training samples)
1
2
3
4
5
6
7
| python train.py --log_name resdcn_all \
--data_dir ~/cyl/Data/PSR_final \
--arch resdcn_101 \
--lr 5e-5 \
--lr_step 40,70 \
--batch_size 16 \
--num_epochs 90 --num_workers 0 --log_interval 5
|
#### Use PSR dataset after remove bg samples (11729 samples)
1
2
3
4
5
6
7
| python train.py --log_name resdcn_all \
--data_dir ~/cyl/Data/PSR_final \
--arch resdcn_101 \
--lr 5e-5 \
--lr_step 40,70 \
--batch_size 16 \
--num_epochs 90 --num_workers 0 --log_interval 5
|
HRNet (↓:4,4,4,4)
Use PSR dataset after remove bg samples (11729 samples)
1
2
3
4
5
6
7
| python train.py --log_name hrnet_all \
--data_dir ~/cyl/Data/PSR_final \
--arch hrnet \
--lr 5e-5 \
--lr_step 40,70 \
--batch_size 16 \
--num_epochs 90 --num_workers 0 --log_interval 5
|
## Hourglass (↓:4,4,4,4)
### Use PSR dataset after remove bg samples (11729 samples)
1
2
3
4
5
6
7
| python train.py --log_name hg_all \
--data_dir ~/cyl/Data/PSR_final \
--arch hourglass_small \
--lr 5e-5 \
--lr_step 40,70 \
--batch_size 16 \
--num_epochs 90 --num_workers 0 --log_interval 5
|
25k dataset
Resdcn+FPNx2 (↓:32,16,8,4)
50 layer
Resnet (pretrained) (↓:32,16,8,4)
