Wgan Introduction __deep

Introduction to Gan the idea of 1.GAN

The GAN model consists of a generative model (generative models) and discriminant models (discriminative model), which is illustrated with an example of generating a picture. Their function is: Generate model G is a generation of image network, it receives a random noise Z, through this noise generated pictures, write G (z). Discriminant Model D is used to determine whether a picture is "real". Its input parameter is the x,x represents a picture, the output D (x) represents the probability that x is the real picture, if is 1, represents 100% is the real picture, but the output is 0, on behalf of cannot be the real picture.

In the training process, the goal of generating network G is to generate real images to deceive the discriminant network D. and D's goal is to try to the G-generated pictures and real pictures separate . In this way, G and D form a dynamic "game process". 2. The problem of the primary Gan is not convergent (non-convergence) is difficult to train
Gradient Extinction (gradient vanishing) collapse problem (mode collapse) model too free uncontrollable reference Goodfellow Ian, Pouget-abadie J, Mirza M, et al. genera tive adversarial nets[c]//advances in neural information processing. 2014:2672-2680. Research progress of Generation against network Gan (II.)--Primary Gan development 1.CGAN (conditional generation against network)

For the problem: the model is too free to be controlled.
Methods: Input more information to the GAN model to learn and produce better samples.
Effect: Improve the quality of the generated image, clearly control some aspects of the image.
Reference: Mirza M, Osindero S. Conditional generative adversarial nets[j]. Computer Science, 2014:2672-2680. Progress of Generation against network Gan (III.)--Conditional Gan 2.DCGAN (deep convolution generation against network)

Combined with supervised learning of CNN and unsupervised Gan
For the problem: Gan training is unstable and often produces meaningless output.
Methods: The generation model and discriminant model were all based on the CNN model, and some changes were made in the structure.
Reference: Radford A, Metz L, Chintala S. Unsupervised representation learning with deep convolutional generative adversarial net WORKS[J]. ARXIV preprint arxiv:1511.06434, 2015. Research progress of Generation against network Gan (v) contribution of --deep convolutional generative adversarial Nerworks,dcgan Wgan 1. Analysis of the difficulties in the training of original Gan 1) Training objectives:

MINGMAXDV (d,g) =ex∼pr[logd (x)]+ez∼pg[1−logd (G (z))] \min \limits_{g}\max\limits_{d}v (d,g) =e_{x\thicksim P_r}[logD (x )]+e_{z\thicksim P_g}[1-logd (g (z))]
x represents the real picture, Z represents the noise of the input G network, and G (z) represents the picture generated by the G network.

Discriminant loss (minimized): −ex∼pr[logd (x)]−ex∼pg[1−logd (x)]-e_{x\thicksim p_r}[logd (x)]-e_{x\thicksim p_g}[1-logd (x)] Builder loss (minimized): EX∼PG[1−LOGD (x)] E_{x\thicksim p_g}[1-logd (x)]
Ex∼pg[−logd (x)] E_{x\thicksim p_g}[-logd (x)] 2) Training process: fixed the generator first, the training discriminant to achieve optimal, and then training generator.
Use the SGD training discriminant to achieve the optimal solution: d∗ (x) =PR (x) Pr (x) +pg (x) d^* (x) =\frac{p_r (x)}{p_r (x) +p_g (x)} Training Builder (when the discriminant is optimal): A.