IRIS Agent

Minimal repo to use the policy of any IRIS agent trained on the 26 Atari 100k games.

Useful Links

• Paper
• Codebase
• Pretrained models

Setup

Install with pip:

pip install git+https://github.com/eloialonso/iris_agent.git

Usage

Create agent:

from iris_agent import Agent

agent = Agent()              # choose from list of games, or
# agent = Agent('Breakout')  # specify game name

Use the policy:

import torch

n = 1
agent.reset(n)
obs = torch.randn(n, 3, 64, 64) # obs is a (n, 3, 64, 64) tensor in [0.,1.], and you should use the standard atari wrappers (see IRIS codebase)
act = agent.act(obs) # act is a (n,) long tensor in {0,...,num_actions-1}