ChunyuanLI / ALICE

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

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ALICE

Adversarially Learned Inference with Conditional Entropy (ALICE)

ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching
Chunyuan Li, Hao Liu, Changyou Chen, Yunchen Pu, Liqun Chen, Ricardo Henao, Lawrence Carin
Duke University. NIPS, 2017.

Alice4Alice: ALICE algorithms for painting the cartoon of Alice's Adventures in Wonderland

Four variants of ALICE on toy datasets

In unsupervised learning case:

In weakly-supervised learning case:

Reproduce figures in the paper

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.

Image Generation Image Reconstruction

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.

Image Generation Image Reconstruction

CelebA

Car2Car

Edge2Shoes

Citation

If you use this code for your research, please cite our paper:

@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}
}

About

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


Languages

Language:Jupyter Notebook 99.7%Language:Python 0.3%