Future-conditioned Unsupervised Pretraining for Decision Transformer

This repository is the official implementation of our ICML 2023 paper Future-conditioned Unsupervised Pretraining for Decision Transformer. Here is the poster.

Requirements

To install requirements, run:

conda env create -f env.yml
conda activate pdt

Data

The D4RL datasets can be downloaded via the following commands:

python data/download_d4rl_gym_datasets.py

Training

To pretrain a model, run this command:

python main.py \
    --data_dir /path/to/data \
    --max_pretrain_iters 50 \
    --num_updates_per_pretrain_iter 1000 \
    --max_online_iters 0 \
    --env hopper-medium-replay-v2 \
    --seed 0

To finetune a pretrained model, run:

python main.py \
    --data_dir /path/to/data \
    --model_path_prefix /path/to/model \
    --model_name model \
    --max_pretrain_iters 0 \
    --online_warmup_samples 10000 \
    --return_warmup_iters 5 \
    --max_online_iters 1500 \
    --num_updates_per_online_iter 300 \
    --env hopper-medium-replay-v2 \
    --seed 0

Evaluation

Run the following script to evaluate a model:

python main.py \
    --eval_only \
    --eval_pretrained \
    --data_dir /path/to/data \
    --model_path_prefix /path/to/model \
    --model_name model \
    --env hopper-medium-replay-v2 \
    --seed 0

Besides, you can also monitor training with Tensorboard:

tensorboard --logdir /path/to/res

Acknowledgement

This repository is based on online-dt, which is licensed under CC-BY-NC. We have made modifications to the models, data processing, and training/evaluation scripts to fit our needs.

Citation

If you use our code or find our work valuable, please cite:

@inproceedings{xie2023future,
  title={Future-conditioned Unsupervised Pretraining for Decision Transformer},
  author={Xie, Zhihui and Lin, Zichuan and Ye, Deheng and Fu, Qiang and Wei, Yang and Li, Shuai},
  booktitle={International Conference on Machine Learning},
  pages={38187--38203},
  year={2023},
  organization={PMLR}
}

fffffarmer/PDT