Is it possible to customize the logger to record some specific parameters?

[zichunxx]

• I have marked all applicable categories:
  • exception-raising bug
  • RL algorithm bug
  • documentation request (i.e. "X is missing from the documentation.")
  • new feature request
• I have visited the source website
• I have searched through the issue tracker for duplicates
• I have mentioned version numbers, operating system and environment, where applicable:

  import tianshou, gymnasium as gym, torch, numpy, sys
  print(tianshou.__version__, gym.__version__, torch.__version__, numpy.__version__, sys.version, sys.platform)

0.5.1 0.29.1 1.12.1 1.24.4 3.9.17 | packaged by conda-forge | (main, Aug 10 2023, 07:02:31) 
[GCC 12.3.0] linux

Hi!

I want to customize a logger to record some desired parameters, like success rate.

I'm trying to get to know the Tianshou API, but I'm not sure if it is easy to implement. Or, are there any examples to be referred to?

I have trained some Mujoco example models. Only these parameters are recorded, such as env_step, length, length_std, test/train reward, etc.

If this is unavailable for the current version of Tianshou? Would it be considered a new feature to record the variation of the success rate during training?

Thanks in advance!

[MischaPanch]

This will likely be addressed by closing #895 and #933. Would you like to help with these? Otherwise, I think I will start working on them quite soon (maybe around end of November), as they are rather high on the priority list

[MischaPanch]

Could you specify what exactly you mean by variation of success rate? @STAY-Melody

[zichunxx]

I want to test the success rate of the current model at certain intervals, e.g., 1e4, during training, which can be achieved by several random test rollouts. In this way, variations in the success rate during training can be recorded. Here is the function that encapsulates what I mean.

[MischaPanch]

We now have dataclasses as returns almost everywhere instead of the previous dicts, so now callbacks can be written. @maxhuettenrauch and I will have a look at this and likely implement soon