search

TEA-Lab / diffusion_reward

[ECCV 2024] 💐Official implementation of the paper "Diffusion Reward: Learning Rewards via Conditional Video Diffusion"
https://diffusion-reward.github.io/
MIT License
70 stars 7 forks source link

About dataset #3

Closed yingchengyang closed 5 months ago

yingchengyang commented 6 months ago

Thanks a lot such a wonderful work! I'm curious about the dataset. First, how can we download the datasets from Hugging Face? Can we directly use the library Datasets? Moreover, if we want to consider other RL tasks with videos, what format should our dataset adopt? It seems that the dataset is like all figures, is it right?

Thanks again!

TaoHuang13 commented 6 months ago

Hi, thanks for your interest in our work.

Please consider this document for downloading our dataset. We suggest using our dataset to reproduce results, though, other datasets could be used depending on your intent.

For dataset format, yes, we only use images. You are encouraged to transform the format of your dataset into one like ours for running code successfully. This will save the time in rewriting the dataloader.

yingchengyang commented 6 months ago

Thanks a lot! It seems that every episode obtains around 30 frames, what is the action repeat during sampling trajectories? Thanks again

TaoHuang13 commented 6 months ago

Hi. Please refer to Table 6 in the Appendix of the paper. We set 2 for Adroit and 3 for MetaWorld.

yingchengyang commented 5 months ago

Thanks a lot!