File size: 4,431 Bytes
03cd173 e0540b6 03cd173 75ae30f 63d6c33 75ae30f 6001dd3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 |
---
license: mit
thumbnail: "https://github.com/conglu1997/v-d4rl/raw/main/figs/envs.png"
tags:
- Reinforcement Learning
- Offline Reinforcement Learning
- Reinforcement Learning from Pixels
- DreamerV2
- DrQ+BC
datasets:
- V-D4RL
---
# V-D4RL
V-D4RL provides pixel-based analogues of the popular D4RL benchmarking tasks, derived from the **`dm_control`** suite, along with natural extensions of two state-of-the-art online pixel-based continuous control algorithms, DrQ-v2 and DreamerV2, to the offline setting. For further details, please see the paper:
**_Challenges and Opportunities in Offline Reinforcement Learning from Visual Observations_**; Cong Lu*, Philip J. Ball*, Tim G. J. Rudner, Jack Parker-Holder, Michael A. Osborne, Yee Whye Teh.
<p align="center">
<a href=https://arxiv.org/abs/2206.04779>View on arXiv</a>
</p>
## Benchmarks
The V-D4RL datasets can be found in this repository under `vd4rl`. They may also be found on [Google Drive](https://drive.google.com/drive/folders/15HpW6nlJexJP5A4ygGk-1plqt9XdcWGI?usp=sharing). **These must be downloaded before running the code.** Assuming the data is stored under `vd4rl_data`, the file structure is:
```
vd4rl_data
└───main
│ └───walker_walk
│ │ └───random
│ │ │ └───64px
│ │ │ └───84px
│ │ └───medium_replay
│ │ │ ...
│ └───cheetah_run
│ │ ...
│ └───humanoid_walk
│ │ ...
└───distracting
│ ...
└───multitask
│ ...
```
## Baselines
### Environment Setup
Requirements are presented in conda environment files named `conda_env.yml` within each folder. The command to create the environment is:
```
conda env create -f conda_env.yml
```
Alternatively, dockerfiles are located under `dockerfiles`, replace `<<USER_ID>>` in the files with your own user ID from the command `id -u`.
### V-D4RL Main Evaluation
Example run commands are given below, given an environment type and dataset identifier:
```
ENVNAME=walker_walk # choice in ['walker_walk', 'cheetah_run', 'humanoid_walk']
TYPE=random # choice in ['random', 'medium_replay', 'medium', 'medium_expert', 'expert']
```
#### Offline DV2
```
python offlinedv2/train_offline.py --configs dmc_vision --task dmc_${ENVNAME} --offline_dir vd4rl_data/main/${ENV_NAME}/${TYPE}/64px --offline_penalty_type meandis --offline_lmbd_cons 10 --seed 0
```
#### DrQ+BC
```
python drqbc/train.py task_name=offline_${ENVNAME}_${TYPE} offline_dir=vd4rl_data/main/${ENV_NAME}/${TYPE}/84px nstep=3 seed=0
```
#### DrQ+CQL
```
python drqbc/train.py task_name=offline_${ENVNAME}_${TYPE} offline_dir=vd4rl_data/main/${ENV_NAME}/${TYPE}/84px algo=cql cql_importance_sample=false min_q_weight=10 seed=0
```
#### BC
```
python drqbc/train.py task_name=offline_${ENVNAME}_${TYPE} offline_dir=vd4rl_data/main/${ENV_NAME}/${TYPE}/84px algo=bc seed=0
```
### Distracted and Multitask Experiments
To run the distracted and multitask experiments, it suffices to change the offline directory passed to the commands above.
## Note on data collection and format
We follow the image sizes and dataset format of each algorithm's native codebase.
The means that Offline DV2 uses `*.npz` files with 64px images to store the offline data, whereas DrQ+BC uses `*.hdf5` with 84px images.
The data collection procedure is detailed in Appendix B of our paper, and we provide conversion scripts in `conversion_scripts`.
For the original SAC policies to generate the data see [here](https://github.com/philipjball/SAC_PyTorch/blob/dmc_branch/train_agent.py).
See [here](https://github.com/philipjball/SAC_PyTorch/blob/dmc_branch/gather_offline_data.py) for distracted/multitask variants.
We used `seed=0` for all data generation.
## Acknowledgements
V-D4RL builds upon many works and open-source codebases in both offline reinforcement learning and online pixel-based continuous control. We would like to particularly thank the authors of:
- [D4RL](https://github.com/rail-berkeley/d4rl)
- [DMControl](https://github.com/deepmind/dm_control)
- [DreamerV2](https://github.com/danijar/dreamerv2)
- [DrQ-v2](https://github.com/facebookresearch/drqv2)
- [LOMPO](https://github.com/rmrafailov/LOMPO)
## Contact
Please contact [Cong Lu](mailto:[email protected]) or [Philip Ball](mailto:[email protected]) for any queries. We welcome any suggestions or contributions!
|