Skip to content

How to reproduce the quality of data ablation study in the paper? #7

Description

@ChandradithyaJ

Hi @ajwagen, thanks a lot for open sourcing the code!

I'm trying to reproduce the results for the ablation study "Does the quality of data πdp is trained on affect DSRL?". In my fork of the repo, I slightly modified the process_robomimic_dataset.py file to allow the creation of train.npz files for data splits from the Robomimic MH datasets. I've also created YAML (pre_diffusion_mlp_worse.yaml) files with the same configurations as in the robomimic can YAMLs to pretrain and eval $\pi_\texttt{dp}$.

However, I always see

[agent.eval.eval_diffusion_agent][INFO] - eval: num episode  120 | success rate   0.000 | avg episode reward  0.000 | avg best reward   0.000

for various training configurations, and surprisingly, data splits (better, okay, worse, worse_okay).

Can you please direct me on how to reproduce those results by providing the required YAML and NPZ files, the $\pi_\texttt{dp}$ checkpoints or clearer guidelines on how to run those experiments?

Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions