mees / calvin

CALVIN - A benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks

Home Page:http://calvin.cs.uni-freiburg.de

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Does it have a Pytorch only version?

wetliu opened this issue · comments

Hello. Thank you for the collecting the datasets. This is awesome.

May I ask if there is an Pytorch only implementation (at least for the dataset usage)?

(Just my own opinion) I feel like as a dataset, using Torch Lightning is very confusing and the configurations are very hard to follow and track (wrapping in the conf folder with many different config files). The dataset itself should be lightweight and simple to use.

Thank you!

The reason for using Lightning DataModule is specifically that it makes it much easier to understand how data preparation works, data splitting and transforms works, instead of normal PyTorch code, where these things are usually scattered across many files. If you want to understand how the Lightning DataModule works I suggest reading https://lightning.ai/docs/pytorch/stable/data/datamodule.html
The underlying dataset is torch DataLoader, so you can also use our DataModule without Lightning: https://lightning.ai/docs/pytorch/stable/data/datamodule.html#datamodules-without-lightning

I created a gist with a minimal example of how to use the dataset without lightning and the nested hydra configs

Hello. Thank you very much for the supportive answers. Deeply appreciate that!