Influence Estimation for Generative Adversarial Networks

This code reproduces the experiments in the following paper:

Naoyuki Terashita, Hiroki Ohashi, Yuichi Nonaka, and Takashi Kanemaru

Influence Estimation for Generative Adversarial Networks

International Conference on Learning Representations (ICLR), 2021.

Requirements

Experiments were conducted on Ubuntu 18.04 with Python 3.6.9 and CUDA 11.2. Other dependencies are summarized in requirements.txt.

Running Experiments

Preparation: extracting 2D-Normal dataset

To make sure the generated 2D-Normal datasets correctly match those of the paper's setting, enter the root of this repository and run,
```
tar -zxvf data/processed_iclr_2d_valid.tar.gz -C .
tar -zxvf data/processed_iclr_2d_cleansing.tar.gz -C .
```

Experiment 1: Estimation Accuracy

Run the following command to reproduce the case of 2D-Normal & FCGAN & Influence on ALL

LUIGI_CONFIG_PARSER=toml LUIGI_CONFIG_PATH=conf/2d_valid.toml python3 main.py TotalizeValid --local-scheduler

Run the following command to reproduce the case of MNIST & DCGAN & Influence on IS / FID

LUIGI_CONFIG_PARSER=toml LUIGI_CONFIG_PATH=conf/mnist_valid.toml python3 main.py TotalizeValid --local-scheduler

Run plot_valid.ipynb to reproduce Figure 1.

Experiment 2: Data Cleansing

Run the following command to reproduce the case of 2D-Normal & FCGAN & Influence on ALL

LUIGI_CONFIG_PARSER=toml LUIGI_CONFIG_PATH=conf/2d_cleansing.toml python3 main.py TotalizeCleansingWrtEval --local-scheduler

Run the following command to reproduce the case of MNIST & DCGAN & Influence on IS / FID

LUIGI_CONFIG_PARSER=toml LUIGI_CONFIG_PATH=conf/mnist_cleansing.toml python3 main.py TotalizeCleansingWrtEval --local-scheduler

Run,
- plot_cleansing.ipynb to reproduce Figure 2.
- plot_quality_2d.ipynb to reproduce Figure 3, Table 9 and 11.
- plot_quality_mnist.ipynb to reproduce Figure 4, Table 10 and 12.

Citation

Please consider citing our paper if it helps your research:

@inproceedings{
terashita2021influence,
title={Influence Estimation for Generative Adversarial Networks},
author={Naoyuki Terashita and Hiroki Ohashi and Yuichi Nonaka and Takashi Kanemaru},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=opHLcXxYTC_}
}

If you have questions, please contact Naoyuki Terashita naoyuki.terashita.sk@hitachi.com