cd data/DfT4D
python preprocess.py --datapath ./ --human bear --pose pose
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/RGBD/ad_RGBD_grad01_lr0001.yaml --mode train --rep sdf --continue_from 1499
CUDA_VISIBLE_DEVICES=0,1,2,3 bash scripts/dist_train.sh 4 --config ./config/RGBD/ad_RGBD_grad01_lr0001.yaml --mode train --rep sdf --batch_size 3 --continue_from 1499
CUDA_VISIBLE_DEVICES=1,2,3 bash scripts/dist_train.sh 3 --config ./config/RGBD/kfusion_RGBD_grad01_lr0001.yaml --mode train --rep sdf --batch_size 10
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/DfT4D/bear/tbase/ad_bear_grad01_lr0001.yaml --mode train --rep sdf --continue_from 4999
Change params: use_sdf_asap_epoch in the yaml file. Tune lr and batch_size.
CUDA_VISIBLE_DEVICES=0,1,2,3 bash scripts/dist_train.sh 4 --config ./config/DfT4D/bear/tbase/ad_bear_grad01_lr0001.yaml --mode train --rep sdf --batch_size 2 --continue_from 4999
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/DfT4D/bear/tbase/ad_bear_grad01_lr0001.yaml --mode train --rep sdf --continue_from 0 --train_from_merge
CUDA_VISIBLE_DEVICES=0,1,2,3 bash scripts/dist_train.sh 4 --config ./config/DfT4D/bear/tbase/ad_bear_grad01_lr0001.yaml --mode train --rep sdf --batch_size 4 --continue_from 0 --train_from_merge
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/RGBD/ad_RGBD_grad01_lr0001.yaml --mode interp --rep sdf --continue_from 4499 --split train
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/RGBD/kfusion_RGBD_grad01_lr0001.yaml --mode interp --rep sdf --continue_from 4499 --split train
If interp_src_fid and interp_tgt_fid are not specified, then by default we interpolate the longest sequence.
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/DfT4D/bear/tbase/ad_bear_grad01_lr0001.yaml --mode interp --rep sdf --continue_from 4499 --split train --interp_src_fid 0 --interp_tgt_fid 1
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/DfT4D/bear/tbase/ad_bear_grad01_lr0001.yaml --mode evaluate --rep sdf --continue_from 9999 --split train
evaluate partial trained model with full trained model
random warp, 比如t = t1, t = t2, 知道他们的translation matrix，我在t1随机sample一些点，得到SDF，然后translaation回t2，再得到一批SDF，尽可能小。
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent