.
└── datasets
    ├── behaviors
    │   └── [task].hdf5 # state-action trajectory datasets for various behaviors.
    └── feedbacks
        └── [task]_[label_type]_[train_or_eval].hdf5 # pairwise trajectory evaluation feedback. [label_type] can be 'syn' for synthetic labels and 'hum' for human feedback labels.

.
├── datasets
│   └── attr_label
│       └── [task]_[label_type].hdf5 # labels given by pre-trained attribute models.
└── results
    ├── attr_func
    │   └── [task]_[label_type].hdf5 # pre-trained transformer-based attribute models.
    ├── diffusion
    │   └── [task]_[label_type].hdf5 # pre-trained diffusion models.
    └── evaluation
        └── [task]_[label_type]_[seed].hdf5 # test logs for pre-trained models.

python train_attr_func.py --task walker --label_type syn  --device [YOUR_DEVICE]
python train_diffusion_model.py --task walker --label_type syn --device [YOUR_DEVICE]
python eval.py --task walker --label_type syn --device [YOUR_DEVICE]
python plot.py --task walker

@inproceedings{dong2024aligndiff,
title={AlignDiff: Aligning Diverse Human Preferences via Behavior-Customisable Diffusion Model},
author={Zibin Dong and Yifu Yuan and Jianye HAO and Fei Ni and Yao Mu and YAN ZHENG and Yujing Hu and Tangjie Lv and Changjie Fan and Zhipeng Hu},
booktitle={The Twelfth International Conference on Learning Representations, {ICLR}},
year={2024},}