Skip to content

Latest commit

 

History

History

models

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

Usage of Trained Agents

This guide explains how to use the pre-trained agents for Atari games using OCAtari. The models must be downloaded and can be used with the provided script from the script folder.

1. Downloading the Models

The trained models can be downloaded from the following link:

Download Models

Ensure that the downloaded models are placed in the appropriate directory, such as OCAtari/models/ or another path of your choosing.

2. Running the Agents

Use the scripts/run_agent.py script to evaluate the pre-trained agents. The script supports specifying the game, model, and configuration options.

Command Structure

python scripts/run_agent.py -g <GAME_NAME> -a <MODEL_PATH> -w <WINDOW_SIZE> -f <FRAMESKIP> -o <OBS_MODE>
  • <GAME_NAME>: Name of the Atari game (e.g., Skiing-v4, Freeway, ALE/Freeway, ALE/Skiing-v5, ...). The ALE in front of the name specifies the newer v5 version.
  • <MODEL_PATH>: Path to the trained model file
  • <WORKERS>: Number of images in the stack (Default is 2 for obj and 4 for dqn-like)
  • <FRAMESKIP>: Number of frames to skip (Default in v5 is 4)
  • <ALGORITHM>: Observation mode (dqn or obj)

Example Commands

2. Running PPO on ALE/Freeway

python scripts/run_agent.py -g ALE/Freeway -a models/Freeway/42/pixel_ppo.cleanrl_model -w 4 -f 4 -o dqn

4. Running Object-Centric PPO on Skiing

python scripts/run_agent.py -g ALE/Skiing -a models/Skiing/42/obj_ppo.cleanrl_model -w 2 -f 4 -o obj

** 3. Use your own models with OCAtari**

  • Check out the run_agent script and replace the load_agent method with your own agent definition
  • Set the parameters to match your model training environment (frameskip, window_size, model)
  • Replace the policy with your predict method or a like

4. Notes

  • Ensure all dependencies are installed (e.g., gymnasium, ale-py, torch).
  • Verify the model paths are correct relative to your script.
  • For headless environments, consider using virtual display tools like Xvfb.
  • More config settings can be found in the script itself, e.g., if you want to record a video of a run.

5. Troubleshooting

Issue: Missing ROMS

  • Check if you install gymnasium with the correct parameters pip install "gymnasium[atari, accept-rom-license]"

Misshape in the neural network

  • Check that you used the correct parameters (-w, -f, -o) for the model you want to use

For further assistance, please reach out to the maintainers or consult the project documentation.