-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathscripts_train_new.sh
15 lines (10 loc) · 2.55 KB
/
scripts_train_new.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# TODO To train to get the improved result for AR-Net(ResNet) (mAP~76.8)
# TODO A. train for new adaptive model
# TODO A-1. prepare each base model (for specific resolution) for 15 epochs
python main_base.py actnet RGB --arch resnet50 --num_segments 16 --gd 20 --lr 0.001 --wd 1e-4 --lr_steps 20 40 --epochs 15 --batch-size 48 --dropout 0.5 --consensus_type=avg --eval-freq=1 --npb --gpus 0 1 2 3 --exp_header actnet_res50_t16_epo15_224_lr.001 --rescale_to 224 -j 36 --data_dir ../../datasets/activity-net-v1.3 --log_dir ../../logs_tsm
python main_base.py actnet RGB --arch resnet34 --num_segments 16 --gd 20 --lr 0.001 --wd 1e-4 --lr_steps 20 40 --epochs 15 --batch-size 48 --dropout 0.5 --consensus_type=avg --eval-freq=1 --npb --gpus 0 1 2 3 --exp_header actnet_res34_t16_epo15_168_lr.001 --rescale_to 168 -j 36 --data_dir ../../datasets/activity-net-v1.3 --log_dir ../../logs_tsm
python main_base.py actnet RGB --arch resnet18 --num_segments 16 --gd 20 --lr 0.001 --wd 1e-4 --lr_steps 20 40 --epochs 15 --batch-size 48 --dropout 0.5 --consensus_type=avg --eval-freq=1 --npb --gpus 0 1 2 3 --exp_header actnet_res18_t16_epo10_112_lr.001 --rescale_to 112 -j 36 --data_dir ../../datasets/activity-net-v1.3 --log_dir ../../logs_tsm
# TODO A-2. joint training for 100 epochs (replace the GGGG with the real datetime shown in your exp dir)
python main_base.py actnet RGB --arch resnet50 --num_segments 16 --lr 0.001 --epochs 100 --batch-size 48 -j 32 --npb --gpus 0 1 2 3 --exp_header jact4_t16_3m124_a.95e.05_ed5_ft15ds_lr.001_gu3_ep100 --ada_reso_skip --policy_backbone mobilenet_v2 --reso_list 224 168 112 84 --backbone_list resnet50 resnet34 resnet18 --skip_list 1 2 4 --accuracy_weight 0.95 --efficency_weight 0.05 --model_paths ../../logs_tsm/GGGG_actnet_res50_t16_epo15_224_lr/models/ckpt.best.pth.tar ../../logs_tsm/GGGG_actnet_res34_t16_epo15_168_lr.001/models/ckpt.best.pth.tar ../../logs_tsm/GGGG_actnet_res18_t16_epo10_112_lr.001/models/ckpt.best.pth.tar --exp_decay --init_tau 5 --policy_also_backbone --policy_input_offset 3 --uniform_loss_weight 3.0 --use_gflops_loss --random_seed 1007 --data_dir ../../datasets/activity-net-v1.3 --log_dir ../../logs_tsm
## TODO B. train for new baseline model (this is also for 100 epochs)
#python main_base.py actnet RGB --arch resnet50 --num_segments 16 --gd 20 --lr 0.001 --wd 1e-4 --lr_steps 100 --epochs 120 --batch-size 48 --dropout 0.5 --consensus_type=avg --eval-freq=1 --npb --gpus 0 1 2 3 --exp_header actnet_tsn_resnet50_seg16_epo120_sz224_b48_lr.001s100 --rescale_to 224 -j 36 --data_dir ../../datasets/activity-net-v1.3 --log_dir ../../logs_tsm