https://github.com/chunyuanli/alice

NIPS 2017: ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching
https://github.com/chunyuanli/alice

bayesian-inference entropy generative-adversarial-network image-translation latent-variable-models

Last synced: 24 days ago
JSON representation

NIPS 2017: ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching

• Host: GitHub
• URL: https://github.com/chunyuanli/alice
• Owner: ChunyuanLI
• Created: 2017-09-05T04:33:25.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2018-08-16T23:52:23.000Z (over 6 years ago)
• Last Synced: 2024-07-30T18:40:23.445Z (5 months ago)
• Topics: bayesian-inference, entropy, generative-adversarial-network, image-translation, latent-variable-models
• Language: Jupyter Notebook
• Homepage:
• Size: 70.1 MB
• Stars: 82
• Watchers: 2
• Forks: 24
• Open Issues: 2
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# ALICE

Adversarially Learned Inference with Conditional Entropy (**ALICE**)

[ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching](https://arxiv.org/abs/1709.01215)  

 [Chunyuan Li](http://chunyuan.li/),

 [Hao Liu](https://hliu96.github.io/), 

 [Changyou Chen](https://www.cse.buffalo.edu/~changyou/), 

 [Yunchen Pu](https://scholar.google.com/citations?user=ftW7RoAAAAAJ&hl=en), 

 [Liqun Chen](https://scholar.google.com/citations?user=T9T8Il0AAAAJ&hl=en), 

 [Ricardo Henao](https://scholar.google.com/citations?user=p_mm4-YAAAAJ),

 [Lawrence Carin](http://people.ee.duke.edu/~lcarin/)  

 Duke University. NIPS, 2017.

![](/plot_generation/figures_alice/alice_log_movie.gif)

*[Alice4Alice](https://github.com/ChunyuanLI/Alice4Alice): ALICE algorithms for painting the cartoon of Alice's Adventures in Wonderland*

## Four variants of ALICE on toy datasets

In *unsupervised learning* case: 

- (a) Explicit cycle-consistency ([`ALICE_l2.py`](/toy_data/ALICE_l2.py)) 

- (b) Implicit cycle-consistency ([`ALICE_A.py`](/toy_data/ALICE_A.py))

In *weakly-supervised learning* case:

- (c) Explicit mapping  ([`ALICE_l2_l2.py`](/toy_data/ALICE_l2_l2.py)) 

- (d) Implicit mapping  ([`ALICE_A_A.py`](/toy_data/ALICE_A_A.py)) 

## Reproduce figures in the paper

[`plot_generation/alice_plots_paper.ipynb`](./plot_generation/alice_plots_paper.ipynb)

## Real datasets

### MNIST

We study the impact of weighting hyperparameter (\lambda) for CE regularizer. The performance of image generation is evaluated by **inception score (ICP)**, and image reconstruction is evaluted by **mean square error (MSE)**.

Best ICP=9.279 ± 0.07, and MSE=0.0803 ± 0.007, when \lambda=1

Note: we pre-trained a "perfect" MNIST classifier (100\% training accuracy) to compute the [inception score for MNIST](https://github.com/ChunyuanLI/MNIST_Inception_Score).

Image Generation             |  Image Reconstruction

:-------------------------:|:-------------------------:

![](/plot_generation/figures/mnist_icp_weighting.png)  |  ![](/plot_generation/figures/mnist_mse_weighting.png)

### CIFAR

Best ICP=6.015 ± 0.0284, and MSE=0.4155 ± 0.2015, when \lambda=1e-6. Larger \lambda leads to lower MSE.

Note: The quality of generated cifar images is evaluated via the [inception score based on ImageNet](https://github.com/openai/improved-gan/tree/master/inception_score).

Image Generation             |  Image Reconstruction

:-------------------------:|:-------------------------:

![](/plot_generation/figures/cifar_icp_weighting.png)  |  ![](/plot_generation/figures/cifar_mse_weighting.png)

### CelebA

### Car2Car

### Edge2Shoes

## Citation

If you use this code for your research, please cite our [paper](https://arxiv.org/abs/1709.01215):

```

@article{li2017alice,

  title={ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching},

  author={Li, Chunyuan and Liu, Hao and Chen, Changyou and Pu, Yunchen and Chen, Liqun and Henao, Ricardo and Carin, Lawrence},

  journal={Neural Information Processing Systems (NIPS)},

  year={2017}

}

```