Skip to content
master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 

GANHopper: Multi-Hop GAN for Unsupervised Image-to-Image Translation

Link for the paper published on Arxiv

News

  • GANHopper was presented at ECCV 2020. See presentation here.

Instrunctions

To run the code, download the datasets into the 'datasets' folder. The internal structure of each dataset should follow the one in datasets/example.

To run the evaluation on this dataset (or on any of the original CycleGAN datasets), please just run

python main.py --dataset_dir={{dataset_dir}}

Where {dataset_dir} is a folder containing a dataset in the format described above. The dog_cat_faces dataset can be downloaded at https://github.com/brownvc/ganimorph

To reproduce the results from the dog_cat_faces dataset please use the default parameters.

To reproduce our results on human2cats, add to the command instruction the parameters --epoch_step=22 and --epoch=22.

Similarly, to reproduce our results on human2dolls, add to the command instruction the parameters --epoch_step=25 and --epoch=25.

The objective function parameters can be controlled with can be set with the following arguments:

           --hybridness: weight of the hybridness loss
           
           --h_hops: total number of translation hops between two domains
           
           --smootheness: weight of the smootheness loss
           
           --L1_lambda: weight on reconstruction loss term between hops in objective
           
           --adversarial: weight of the adversarial loss

Dependencies

tensorflow-gpu=1.9.0
numpy=1.15.2
scipy=1.1.0
pillow=3.3.0
imageio=2.4.1

About

No description, website, or topics provided.

Resources

Releases

No releases published

Packages

No packages published

Languages