Skip to content

This is the pytorch implementation of ICML 2018 paper - Self-Imitation Learning.

License

Notifications You must be signed in to change notification settings

TianhongDai/self-imitation-learning-pytorch

Repository files navigation

Self-Imitation-Learning with A2C

MIT License
This is the pytorch version of the A2C + SIL - which is basiclly the same as the openai baselines. The paper could be found Here.

TODO List

  • Add PPO with SIL
  • Add more results

Requirements

  • python-3.5.2
  • openai-baselines
  • pytorch-0.4.0

Installation

Install OpenAI Baselines (Need to use the previous version of openai-baselines, will solve in the future.)

# clone the openai baselines
git clone https://github.com/openai/baselines.git
cd baselines
git checkout 366f486
pip install -e .

How to use the code

Train the network:

python train.py --env-name 'PongNoFrameskip-v4' --cuda (if you have the GPU)

Test the network:

python demo.py --env-name 'PongNoFrameskip-v4'

You could also try the A2C algorithm without SIL by adding flag --no-sil:

python train.py --env-name 'PongNoFrameskip-v4' --cuda --no-sil

Training Performance

Because of time, I just run Pong with 2 million steps. The results of MontezumaRevenge will be uploaded later! Scheme
Another results for the Freeway which is correspond with the original paper.
freeway

Demo: FreewayNoFrameskip-v4

freewaydemo

Acknowledgement

@junhyukoh for original code

About

This is the pytorch implementation of ICML 2018 paper - Self-Imitation Learning.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages