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
Top: DiscoGAN Middle: NAM: Bottom: Source
Top: DiscoGAN Middle: NAM: Bottom: Source
- Download Edges2Shoes data:
cd data
sh get_data.py
cd ..
- Train DCGAN unconditional generative model for the A domain:
cd code
python train_gen.py
- Use NAM to train a mapping from A to B:
python train_nam.py
- 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.
This project is CC-BY-NC-licensed.