search

zzq-bot / offline-marl-framework-offpymarl

Benchmarked implementations of Offline Multi-Agent RL Algorithms based on PyMARL codebase.
21 stars 1 forks source link

Integrate OG-MARL Datasets #1

Open jcformanek opened 4 months ago

jcformanek commented 4 months ago

Hi there, I came across your project recently. Its looking very good. I wanted to know if you would be interested in integrating OG-MARL datasets with your framework.

https://github.com/instadeepai/og-marl

We would really appreciate you benchmarking your algorithms on our datasets and I am happy to help.

zzq-bot commented 4 months ago

Hi, sorry for the delayed response. Thank you for your suggestion. I am willing to benchmark the algorithms on OG-MARL. However, I have some issues to deal with in the next two days. I will start preparing the OG-MARL dataset and begin testing next week. If I need any assistance, I will reach out to you. 🎉

zzq-bot commented 4 months ago

@jcformanek Hi, I was trying to download the dataset provided by OG-MARL following examples/dataset_api_demo.ipynb/download_dataset.py; but I came cross some problems:

  1. Due to regional reasons, the dataset download speed is very slow, and it takes me more than 24 hours to download the 3m .zip. Could you please upload it to Google Drive? It would be even better if it could be converted into an h5 file.
  2. a minor suggestion: the parameter may be mistaken in download_dataset.py; FLAGS.env_name, FLAGS.scenario_name -> FLAGS.env, FLAGS.scenario
jcformanek commented 4 months ago

Thanks for your reply, I am pleased that you are keen to try the OG-MARL datasets.

I am really glad you brought to my attention the regional download issue. I will try to work on fixing that. But for now, here is a google drive link to a subset of the datasets. Let me know if you want any other ones. I have only uploaded the SMAC datasets for now. https://drive.google.com/drive/folders/1lw-e5VwIdCtmsGWgQG902yZRArU69TrH?usp=sharing

I would prefer to use our Flashbax vault API instead of h5 files. I think you will find it easy enough to work with. Converting Vaults to numpy arrays should be easy enough if you follow examples/dataset_api_demo.ipynb. But if you have any issues please do just ask and I will be happy to help.

zzq-bot commented 4 months ago

Thank you for your help, I have successfully downloaded the dataset from the Google Drive URL you provided😁. I am willing to get used to Flashbax vault API through examples/dataset_api_demo.ipynb and then start the benchmarking.

jcformanek commented 1 month ago

Hi there, how is the benchmarking going? I saw on the README that your systems now work with OG-MARL datasets, that is great 🥳

zzq-bot commented 3 weeks ago

Sorry for the delayed response. I had completed the benchmarking on the 3m datasets earlier, but I got busy with some competitions and my internship afterwards, so I haven't had the time to continue the benchmarking😭. I’ll get back to it as soon as I can.