[Question] DQN save model taking up large hard disk space

[WreckItTim]

❓ Question

Hello,

I am saving my DQN model to file and noticed it is taking up about 30 GB of disk space. When I unzip the folder, I noticed that over 28 GB is coming from the "data" object within the zip folder. This is listed on the SB3 docs as: "a JSON file of class-parameters (dictionary)". The actual pytorch model is under 1 GB. It seems like something is being serialized in my implementation that shouldn't be. This is about the same size as the replay buffer since I am using DRL with images. I noticed that when I train the same code with a different observation space (some small vectors instead of images) that the size of the zip file goes back to below a gig. So it seems related to the observation space some how. The file is too big to open.

Checklist

• I have checked that there is no similar issue in the repo
• I have read the documentation
• If code there is, it is minimal and working
• If code there is, it is formatted using the markdown code blocks for both code and stack traces.

[WreckItTim]

This was because I was overwriting the exploration_schedule function with a custom one, which was a custom* class method that also had the sb3model (and accordingly replay buffer) as attributes. So it was writing the replay buffer during serialization. Oops!

I am going to pull the custom exploration_schedule method out of the class method, but it requires some additional arguments other than just current_progress_remaining as input to the method call. Is there a more SB3 native way for using a custom exploration schedule with additional custom arguments?

Thanks!

[araffin]

Is there a more SB3 native way for using a custom exploration schedule with additional custom arguments?

for additional args, you will need to subclass DQN.