UniHCP: A Unified Model for Human-Centric Perceptions
Usage
Preparation
- Install all required dependencies in requirements.txt.
- Replace all
path...to...
in the .yaml configuration files to the absolute path to corresponding dataset locations. - Place MAE pretrained weight mae_pretrain_vit_base.pth under
core\models\backbones\pretrain_weights
folder.
*Only slurm-based distributed training & single-gpu testing is implemented in this repo.
Experiments
All experiment configurations files and launch scripts are located in experiments/unihcp/release
folder.
To perform full multi-task training for UniHCP, replace <your partition>
in train.sh
launch script and run:
sh train.sh 88 coslr1e3_104k_b4324g88_h256_I2k_1_10_001_2I_fairscale_m256
To perform evaluations, keep the test_info_list assignments corresponding to the tests you want to perform
, replace <your partition>
, then run :
sh batch_test.sh 1 coslr1e3_104k_b4324g88_h256_I2k_1_10_001_2I_fairscale_m256
Note that in this case, the program would look for checkpoints located at experiments/unihcp/release/checkpoints/coslr1e3_104k_b4324g88_h256_I2k_1_10_001_2I_fairscale_m256
Pretrained Models
Please send the signed agreement to [email protected]
to get the download link.