This repo provides an unofficial PyTorch Implementation of Spectral Normalization for Generative Adversarial Networks (SNGAN)1 with specialization in Anime faces2 generation.
-
original SNGAN
- differentiable power iteration
-
$\gamma$ -reparameterization - ResNet architecture
- hinge loss 3
-
dataset-specific tuning
- set
$n_{dis} = 1$ - use 1x1 conv instead of 3x3 conv for penultimate layer of G
- replace
avgpool2ddown-sampling with stride-2 conv
- set
-
TTUR4
-
mixed precision (-11% time)5
-
cuDNN benchmark
-
FID metrics + precomputed statistics- InceptionV3 pretrained on ImageNet is not suitable for anime data
- data-specific feature extractor is needed
-
exponential moving average
-
visualizations 😄
evolving over 50 training epochs
from top to bottom: the variance varies from 0.3 to 3 cubically
Footnotes
-
Miyato, Takeru, et al. "Spectral Normalization for Generative Adversarial Networks." International Conference on Learning Representations. 2018. ↩
-
Chao, Brian. ‘Anime Face Dataset: a collection of high-quality anime faces’. GitHub, https://github.com/bchao1/Anime-Face-Dataset. ↩
-
Lim, Jae Hyun, and Jong Chul Ye. "Geometric gan." arXiv preprint arXiv:1705.02894 (2017). ↩
-
Heusel, Martin, et al. "Gans trained by a two time-scale update rule converge to a local nash equilibrium." Advances in neural information processing systems 30 (2017). ↩
-
"Automatic Mixed Precision For Deep Learning". NVIDIA Developer, https://developer.nvidia.com/automatic-mixed-precision. ↩
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent