/NAM

Non-Adversarial Unsupervised Domain Mapping

Primary LanguagePythonOtherNOASSERTION

NAM: Non-Adversarial Mapping

This repo contains PyTorch code replicating the main ideas presented in:

  • NAM - Unsupervised Cross-Domain Image Mapping without Cycles or GANs
    Yedid Hoshen and Lior Wolf, ICLR 2018 Workshop
    ICLR Manuscript

  • Non-Adversarial Unsupervised Domain Mapping
    Yedid Hoshen and Lior Wolf, ECCV 2018
    https://arxiv.org/abs/1806.00804

Examples

Edges2Bags

Alt text
Top: DiscoGAN Middle: NAM: Bottom: Source

Edges2Shoes

Alt text
Top: DiscoGAN Middle: NAM: Bottom: Source

Variation in Outputs:

Alt text
Alt text

Getting Started

  1. Download Edges2Shoes data:
cd data
sh get_data.py
cd ..
  1. Train DCGAN unconditional generative model for the A domain:
cd code
python train_gen.py
  1. Use NAM to train a mapping from A to B:
python train_nam.py
  1. Evaluate multiple image analogies:
python eval_variation.py $image_id

Where $image_id is replaced with the ID of the image you wish to map.

Note: DCGAN training can diverge sometimes. Unconditional samples from each epoch are available in "code/unconditional_ims/". If DCGAN training diverged, simply re-run it.

License

This project is CC-BY-NC-licensed.