高分辨率256*256人脸生成效果介绍及代码

What is this?

这些人脸都是生成的，不是真实的，生成的人脸分辨率在256*256（分辨率很高了）

理论上现在可以生成无穷个人脸

Note: 神经网络仍在收敛中

What is a GAN?

GAN包括生成网络和辨识网络，他们共同训练但可以单独使用。参考paper如下

More detail

理论上可以生成无限的脸，但实际上还有一些坑，还需要注意一些：

How?

众所周知，训练GAN非常困难. In order to train at 256 x 256 we utilize:

• Hyperchamber.io, 超参数搜索的训练网络的服务
• A 定制 Discriminator and Generator
• A 定制 Vae(variational encoder), jointly trained
• Custom GAN training technique 训练方法也定制
• GPU：Nvidia Titan X
• 训练时间约一周

代码开源(还在改进). Our training data was custom built using dlib to identify facial landscape points, then rotate and crop at a certain width/height. In total, this network was trained on 4万张人脸 human female faces.

Links

• Hyperchamber -
• hypergan - The network trainer/runner - very alpha
• The configuration for this network is here:https://hyperchamber.255bits.com/ba3051e944601e2d98b601f3347df0b1/40k_overfit_3:1.2/samples/9602b13bb5669064d636f88b144d9067
• The dataset was independently created. If you need it, email me.

HyperGAN is an open implementation 很多不同类型的 GANs (generative adversarial networks). It is currently in open alpha as it relies on Hyperchamber.

GANs are known for being hard to train. HyperGAN has three unique features:

• Runs on a directory of images
• Searches for a good network configuration (using Hyperchamber) 搜索网络配置进行超参数设置
• Has many recent advancements 超前

Each GAN trained will learn different aspects of your data. Many GANs wont work at all. Some will converge to a few examples and not establish a meaningful feature space. There are many many ways for a GAN to fail. GAN训练失败有很多原因

HyperGAN on github https://github.com/255BITS/HyperGAN

focused on scalability and ease-of-use. 关注扩展性和易用。

Features

• Efficient GAN implementation
• Semi-supervised or unsupervised learning(works with and without labels)
• Variational methods
• InfoGAN-inspired categories
• Minibatch normalization
• Adversarial inference
• Flask server mode

Thanks to

• GAN paper - https://arxiv.org/abs/1406.2661
• DCGAN paper - https://arxiv.org/abs/1511.06434
• Adversarial inference - https://arxiv.org/abs/1606.00704
• Improved GAN - https://arxiv.org/abs/1606.03498
• Tensorflow - https://www.tensorflow.org/

Related work

• dcgan-tensorflow - https://github.com/carpedm20/DCGAN-tensorflow
• openai improved-gan - https://github.com/openai/improved-gan
• openai infoGAN - https://github.com/openai/InfoGAN
• VisioFaces - http://www.faculty.idc.ac.il/arik/seminar2009/papers/VisioFaces.pdf