https://github.com/modanesh/differential_ig

Source code for the differential saliency method used in "Re-understanding Finite-State Representations of Recurrent Policy Networks"
https://github.com/modanesh/differential_ig

computer-vision explainable-ai pytorch reinforcement-learning saliency

Last synced: 6 months ago
JSON representation

Source code for the differential saliency method used in "Re-understanding Finite-State Representations of Recurrent Policy Networks"

• Host: GitHub
• URL: https://github.com/modanesh/differential_ig
• Owner: modanesh
• Created: 2020-03-16T16:36:51.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2023-10-04T01:05:33.000Z (about 2 years ago)
• Last Synced: 2025-04-10T21:53:00.708Z (6 months ago)
• Topics: computer-vision, explainable-ai, pytorch, reinforcement-learning, saliency
• Language: Python
• Homepage: https://arxiv.org/abs/2006.03745
• Size: 5.11 MB
• Stars: 11
• Watchers: 1
• Forks: 1
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Differential Integrated Gradient

This is the implementation of the differential saliency method used in "[Re-understanding Finite-State Representations of Recurrent Policy Networks](https://arxiv.org/abs/2006.03745)", accepted to the **International Conference on Machine Learning (ICML) 2021**. 

## Installation

* Python 3.5+

* To install dependencies:

    ```bash

    pip install -r requirements.txt

    ```

## Usage

You can use ```main_IG.py``` or ```main_IG_control.py``` for experimenting with Atari and Control Tasks from OpenAI Gym.

To begin, you  need to load and use models trained here: [MMN](https://github.com/koulanurag/mmn). Once you took all the steps, you end up with a MMN model, and that's what is needed in this repo. Trained models should be put into the ```inputs``` directory with a proper name.

Having the models, it's time to run the code. To do that, just run the following command to get the results for Atari games:

```

python main_IG.py --env_type=atari --input_index=43 --baseline_index=103 --env PongDeterministic-v4 --qbn_sizes 64 100 --gru_size 32

```

Values of the input arguments can be changed according to your interest.

And the following command to get the results for control tasks:

```

python main_IG_control.py --env_type=classic_control --input_index=10 --baseline_index=106 --env CartPole-v1 --qbn_sizes 4 4 --gru_size 32

```

Results will be saved into the ```results``` folder. In the repo, we have already provided sample results. For example, in the case of CartPole, an output will look like the following:



  



## Citation

If you find it useful in your research, please cite it with:

```

@inproceedings{danesh2021re,

  title={Re-understanding Finite-State Representations of Recurrent Policy Networks},

  author={Danesh, Mohamad H and Koul, Anurag and Fern, Alan and Khorram, Saeed},

  booktitle={International Conference on Machine Learning},

  pages={2388--2397},

  year={2021},

  organization={PMLR}

}

```