Awesome

pytorch-MNIST-CelebA-cGAN-cDCGAN

Pytorch implementation of conditional Generative Adversarial Networks (cGAN) [1] and conditional Generative Adversarial Networks (cDCGAN) for MNIST [2] and CelebA [3] datasets.

The network architecture (number of layer, layer size and activation function etc.) of this code differs from the paper.
CelebA dataset used gender lable as condition.
If you want to train using cropped CelebA dataset, you have to change isCrop = False to isCrop = True.
you can download
- MNIST dataset: http://yann.lecun.com/exdb/mnist/
- CelebA dataset: http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

Implementation details

cGAN

GAN

cDCGAN

Loss

Resutls

MNIST

Generate using fixed noise (fixed_z_)

<table align='center'> <tr align='center'> <td> cGAN</td> <td> cDCGAN</td> </tr> <tr> <td><img src = 'MNIST_cGAN_results/generation_animation.gif'> <td><img src = 'MNIST_cDCGAN_results/MNIST_cDCGAN_generation_animation.gif'> </tr> </table>

MNIST vs Generated images

<table align='center'> <tr align='center'> <td> MNIST </td> <td> cGAN after 50 epochs </td> <td> cDCGAN after 20 epochs </td> </tr> <tr> <td><img src = 'MNIST_cGAN_results/raw_MNIST.png'> <td><img src = 'MNIST_cGAN_results/MNIST_cGAN_50.png'> <td><img src = 'MNIST_cDCGAN_results/MNIST_cDCGAN_20.png'> </tr> </table>

Learning Time
- MNIST cGAN - Avg. per epoch: 9.13 sec; Total 50 epochs: 937.06 sec
- MNIST cDCGAN - Avg. per epoch: 47.16 sec; Total 20 epochs: 1024.26 sec

CelebA

Generate using fixed noise (fixed_z_; odd line - female (y: 0) & even line - male (y: 1); each two lines have the same style (1-2) & (3-4).)

<table align='center'> <tr align='center'> <td> cDCGAN</td> <td> cDCGAN crop</td> </tr> <tr> <td><img src = 'CelebA_cDCGAN_results/CelebA_cDCGAN_generation_animation.gif'> <td><img src = 'CelebA_cDCGAN_crop_results/CelebA_cDCGAN_crop_generation_animation.gif'> </tr> </table>

CelebA vs Generated images

<table align='center'> <tr align='center'> <td> CelebA </td> <td> cDCGAN after 20 epochs </td> <td> cDCGAN crop after 30 epochs </td> </tr> <tr> <td><img src = 'CelebA_cDCGAN_results/raw_CelebA.png'> <td><img src = 'CelebA_cDCGAN_results/CelebA_cDCGAN_20.png'> <td><img src = 'CelebA_cDCGAN_crop_results/CelebA_cDCGAN_crop_30.png'> </tr> </table>

CelebA cDCGAN morphing (noise interpolation)

<table align='center'> <tr align='center'> <td> cDCGAN </td> <td> cDCGAN crop </td> </tr> <tr> <td><img src = 'CelebA_cDCGAN_results/CelebA_cDCGAN_morp.png'> <td><img src = 'CelebA_cDCGAN_crop_results/CelebA_cDCGAN_crop_morp.png'> </tr> </table>

Learning Time
- CelebA cDCGAN - Avg. per epoch: 826.69 sec; total 20 epochs ptime: 16564.10 sec

Development Environment

Ubuntu 14.04 LTS
NVIDIA GTX 1080 ti
cuda 8.0
Python 2.7.6
pytorch 0.1.12
torchvision 0.1.8
matplotlib 1.3.1
imageio 2.2.0

Reference

[1] Mirza, Mehdi, and Simon Osindero. "Conditional generative adversarial nets." arXiv preprint arXiv:1411.1784 (2014).

(Full paper: https://arxiv.org/pdf/1411.1784.pdf)

[2] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. "Gradient-based learning applied to document recognition." Proceedings of the IEEE, 86(11):2278-2324, November 1998.

[3] Liu, Ziwei, et al. "Deep learning face attributes in the wild." Proceedings of the IEEE International Conference on Computer Vision. 2015.