MT-GAN-PyTorch

PyTorch Implementation of AAAI-2020 Paper "Learning to Transfer: Unsupervised Domain Translation via Meta-Learning"

Dependency:

Python 3.6

PyTorch 0.4.0

Usage:

Unsupervised Domain Translation via Meta-Learning

1. Downloading labels2photos, horses2zebras, summer2winter, apple2orange, monet2photo, cezanne2photo, ukiyoe2photo, vangogh2photo, photos2maps and labels2facades datasets following CycleGAN.

2. Organize these translation datasets as:

   meta_datarooot
   ├── vangogh2photo
   |   ├── trainA
   |   ├── trainB
   |   
   ├── ukiyoe2photo
       ├── trainA
       ├── trainB
   ...
   

3. Train MT-GAN on 10-shot tranlation:

   $ python train.py --name mtgan_results --model mt_gan --meta_dataroot meta_datarooot --k_spt 10 --k_qry 10 --finetune_step 1000

4. Test MT-GAN on 10-shot translation:

   $ python test.py --name mtgan_results --model mt_gan --meta_dataroot meta_datarooot --k_spt 10 --k_qry 10 --finetune_step 1000

5. Train MT-GAN on 5-shot translation:

   $ python train.py --name mtgan_results --model mt_gan --meta_dataroot meta_datarooot --k_spt 5 --k_qry 5 --finetune_step 1000

6. Test MT-GAN on 5-shot translation:

   $ python test.py --name mtgan_results --model mt_gan --meta_dataroot meta_datarooot --k_spt 5 --k_qry 5 --finetune_step 1000