Create a separate repo for Wrappers ?

[zuoxingdong]

Because there are quite some Issues/PRs for new features which could be done through implementing a Wrapper, plus there are very useful wrappers in baselines repo for RL algorithms, e.g. Atari preprocessing/DQN, continuous action clipping, normalizing observations, scaling rewards etc.

Therefore, will it be a good idea to have another repo maintained by OpenAI for keeping a wide collections of Wrappers in one place (+ unit tests), this makes gym itself lightweight and helpful for the community to access to standardized, well-tested wrappers.

What do you think @pzhokhov ?

[pzhokhov]

Having a collection of standardized, well-tested wrappers - sounds very useful, I agree with that. I am not sure separate repo is justified though; rather we can put them (and the unit tests) into the gym. Per discussion with @christopherhesse, env-specific standard wrappers could live in the gym, whereas wrappers that are algorithm-specific/related can live closer to algorithms, i.e. in baselines. Let's keep this issue open until we complete migration of wrappers into gym and their testing.

[zuoxingdong]

Thanks @pzhokhov , that sounds exciting ! I've made a small list here for further discussion:

• Standard

  • [x] ClipAction: clip the received action with upper and lower bounds. Valid for Box action space
  • [x] SignReward: apply sign function to bin the reward
  • [x] ClipReward: clip the raw reward with upper and lower bounds.
  • [x] ScaleReward: rescale the reward by a factor
  • [ ] WarpFrame: downsample the image observation. Note: maybe better name with ResizeFrame/DownsampleFrame
  • [x] FlattenObservation
  • [x] GrayScaleObservation: convert pixel observation to gray scale
  • [x] TimeAwareObservation: append time step to the observation. Refer to paper, Time Limits in Reinforcement Learning

• Standardization: better applies to VecEnv

  • [ ] StandardizeObservation: standardize the observation by online estimation of moments
  • [ ] StandardizeReward: standardize the reward by online estimation of variance. Note: NOT subtracted by mean
  • [ ] VecMonitor: record episodic length, returns and elapsed time.

• Specific

  • [x] NoopResetEnv: atari
  • [x] FireResetEnv: atari
  • [x] EpisodicLifeEnv: atari
  • [x] MaxAndSkipEnv: atari
  • [x] FrameStack: DQN
  • [x] ScaledFloatFrame: DQN
  • [x] LazyFrames: DQN

[zuoxingdong]

@pzhokhov I'd be happy to implement some as the discussion leads to migrate that wrapper.

[AdilZouitine]

Hi @zuoxingdong , I saw you did a lot of PRs implementing the wrappers you mentioned!

My question is if there is still some wrapper left to implement, I would be happy to help you :smile:

Thank you.

[zuoxingdong]

Hi @AdilZouitine , thanks a lot for your interest ! I've modified a checklist above, for now, probably the standardization for observation and reward in vectorized environment are missing. Would you like to implement them ?

[AdilZouitine]

Hi @zuoxingdong, yes, one of them interests me, it will be an opportunity to learn more about environment vectors.

For my internship, I had to implement a wrapper similar to the WarpFrame that you propose in your list. I can also take care of this one.

Best regards :smile:

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.