elicassion/StARformer

Difference between the two starformer.py files?

Closed this issue · 4 comments

shahaamirbader commented

Hi,

Congratulations for such an awesome work. I being new to DT, was wondering what's the main differences between the two starformer.py files in each environment? And also what is the need to have two different files for each environment. Why not one single straformer.py suffice for both?

elicassion commented

Hi @shahaamirbader,

The main difference is input. The starformer.py under atari_and_dmc deals with image inputs, while the one under gym deals with numerial state inputs.

Our paper does not include the results on gym environments, but StARformer does work for numerical state inputs and we observed improvements there.

shahaamirbader commented

Thanks for clarification. This brings me to my next question. Can your code be used as a code base for other image related inputs for a custom environment?

elicassion commented

Sure. Like what we did in atari_and_dmc/run_star_atari.py and atari_and_dmc/run_star_dmc.py, you will need to modify:

  • Environment
  • Dataset
  • Input & output dimension / space

I think taking a look into those two files will give you enough hints.

shahaamirbader commented

thanks a ton. closing this.

  • 👍1