GAN, DCGAN, cGAN Paper Reading

GAN: Generative Adversarial Nets

GAN跟传统神经网络不一样的地方就是它的loss函数，一个神经网络有很多个部件不稀奇，但是train的时候是max-min的loss是很创新的事情。简单介绍，GAN有一个Generator（G），一个Discriminator（D）。G生成，D来区分这个生成的结果是像真的呢，还是像假的呢。所以D其实相当于解决一个二分类问题。当然G生成的越真实越好是我们的目标。

所以整个GAN的主要的任务其实是生成，Generator的output是我们想要的结果。但是这个结果是通过discriminator的训练来进行指导的。所以G和D呢，你中有我，我中有你。在对抗中互相提高。所以如果想进步，挑选一个厉害的竞争对手哦，哈哈。

1. traditional GAN 2014

1.1 Paper: 

https://arxiv.org/pdf/1406.2661

1.2 Understanding:

1. E.g. From noise to real image.
2. Generator: G(z;θg)
   1. Input: noise z;
   2. Output: generated image G(z).
3. Discriminator: D(x; θd)
   1. Input: generated image: G(z), label: (0) + real image: x, label: (1)
   2. Output: predicted probability
4. Loss:
   1. Log(D(x))+log(1-D(G(z)))
   2. For D: minimize term 1; for G: minimize term 2.

1.3 Pic: 

1.4 Network:

Simple dense layers 

2. Conditional GAN 2014

2.1 Paper: (unpaired cGAN)

https://arxiv.org/pdf/1511.06434.pdf

2.2 Key improvement:

add a condition y, y can be image, text or sound info.....

2.3 Type: (pair means the input data dependence)

1. Un-paired cGAN
2. Paired cGAN

2.4 Understanding:

1. E.g. MNIST with condition y(label from 0-9)----unpaired cGAN
2. Generator: G(z, y; θg)
   1. Input: noise z (100), label y(10)----one-hot label
      1. Map to hidden layer with Relu respectively, 200 and 1000 units
      2. Concatenate these two layers into 1200 units as the input of thefollowing layers
   2. Output: 784 units. (Mnist pic is 28*28)
3. Discriminator: D(x; θd)
   1. Input: (real image (x), label y) 1, (generated image G(x,y) ) 0
      1. X and y map to maxout layer, 240 units and 5 pieces, 50 units and 5 pieces.
      2. Both map to hidden layer, 240 units and 4pieces and concatenate to feed sigmoid
   2. Output: predicted probability
4. Loss: Log(D(x, y))+log(1-D(G(z, y)))
   1. For D: minimize term 1; for G: minimize term 2.

2.5 Pic:

 

2.6 Network:

Simple dense layers

2.7 there are some papers about paired cGAN:

----This part I'll implement later.....

----To be continued!!!!

3.DCGAN: deep convolutional GAN 2015

3.1 Paper: 

arxiv.org/abs/1411.1784

3.2 Key improvement:

Combine CNN and traditional GAN, the layers of GAN are convolutional layers instead of dense layers

3.3 Key ideas:

1. Don't use MAX pooling or average pooling, while doing downsampling, change the stride of CNN
2. Don't use fully connected layers, use reshape to change the input to larger channels.
3. Use Batch normalization layer after CNN and activation layer.

3.4 Architecture pic

3.5 Training details:

1. Pixels are maped b.t.w(between) [-1, 1], which is corresponding to the tanh activation function.
2. SGD，mini batch = 128;
3. weight initialization, mean=0, standard=0.02；
4. LeakyRelu, slope of the leak = 0.2;
5. Adam optimizer, learning rate = 0.0002, momentum, beta_1=0.5。