Pytorch implementation of CycleGAN [1].
- apple2orange
- apple training images: 995, orange training images: 1,019, apple test images: 266, orange test images: 248
- horse2zebra
- horse training images: 1,067, zebra training images: 1,334, horse test images: 120, zebra test images: 140
apple2orange (after 200 epochs)
Input |
Output |
Reconstruction |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Input |
Output |
Reconstruction |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
- Learning Time
- apple2orange - Avg. per epoch: 299.38 sec; Total 200 epochs: 62,225.33 sec
horse2zebra (after 200 epochs)
Input |
Output |
Reconstruction |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Input |
Output |
Reconstruction |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
- Learning Time
- horse2zebra - Avg. per epoch: 299.25 sec; Total 200 epochs: 61,221.27 sec
- Ubuntu 14.04 LTS
- NVIDIA GTX 1080 ti
- cuda 8.0
- Python 2.7.6
- pytorch 0.1.12
- matplotlib 1.3.1
- scipy 0.19.1
[1] Zhu, Jun-Yan, et al. "Unpaired image-to-image translation using cycle-consistent adversarial networks." arXiv preprint arXiv:1703.10593 (2017).
(Full paper: https://arxiv.org/pdf/1703.10593.pdf)