Research progress of generative anti-network gan (III.)--Condition Gan

Preface
     This article first introduces the build model and then focuses on the resulting model (generative Models), the research and development of the generative adversarial network is generated. According to Gan main thesis, gan applied paper and gan related papers, the author sorted out 45 papers in recent two years, focused on combing the links and differences between the main papers, and revealing the research context of the generative antagonism network.
The papers covered in this article are:
[1] Goodfellow Ian, Pouget-abadie J, Mirza M, et al. generative adversarial nets[c]// Advances in neural information processing Systems. 2014:2672-2680.
[2] Mirza M, Osindero S. Conditional generative adversarial nets[j]. Computer Science, 2014:2672-2680.
[3] Denton E L, Chintala S, Fergus R. Deep generative Image Models using a Laplacian Pyramid of adversarial networks[ C]//advances in neural information processing systems. 2015:1486-1494. 3. Conditional generation anti-network, Conditional generative adversarial Networks 3.1 Cgan Ideas

     development of anti-network Gan Research (II.)--The original Gan suggests that, compared with other generative models, Gan is no longer required to have a hypothetical data distribution, that is, formulate p (x) is not required. Instead of using a distribution to directly sample the sampling, the real data can be completely approximated theoretically, which is the biggest advantage of Gan. However, this method does not require pre-modeling the disadvantage is too free, for larger pictures, more pixel situation, based on simple GAN way is not very controllable. In order to solve the problem of Gan too free, a very natural idea is to add some constraints to Gan, so there is conditional generative adversarial Nets (Cgan) "Mirza M, Osindero S. Conditional ". This work proposes a conditionally constrained Gan, which introduces condition variable y (conditional variable y) in the modeling of the Generation model (D) and discriminant model (G), and adds the condition to the model with additional information Y. Can guide the data generation process. These condition variable y can be based on a variety of information, such as category labels, for image repair part of the data [2], from different modes (modality) data. If the condition variable y is a category label, it can be seen as Cgan is an improvement of the purely unsupervised GAN into a supervised model. This simple and straightforward improvement is proven to be very effective and widely used in subsequent work [3,4]. The work of Mehdi Mirza et al. is to generate an image of the specified category on the Mnist DataSet with the category label as the condition variable. The author also explores the application of Cgan in multi-modal learning for automatic image labeling, and on the Mir Flickr25000 dataset, the vector of the tag of the image is generated by the image feature as the conditional variable. 3.2 Conditional generative adversarial Nets 3.2.1 Generative adversarial Nets

Generative adversarial nets is a new method of training generative model proposed by GOODFELLOW[5, which contains two models of "confrontation": The Generation Model (G) is used to capture data distribution, Discriminant model (D) is used to estimate the probability that a sample comes from a real data rather than a sample generation. In order to learn the generation distribution on the real DataSet X PG, the model G builds a mapping function g (Z;ΘG) from the prior distribution of the Pz (z) to the data space. The input of the discriminant model D is a real image or generates an image, and D (x;θd) outputs a scalar that indicates the probability of the input sample coming from the training sample rather than generating the sample.
The model G and D are trained simultaneously: fixed discriminant model D, which adjusts the parameters of G to minimize the expectation of log (1−d (g (z)), the fixed generation model G, and the parameter of D to maximize the expectation of LOGD (X) + log (1−d (g (z))). This optimization process can be attributed to a "two-dollar Minimax game (minimax two-player game)" Problem:

3.2.2 Conditional adversarial Nets

The Conditional Generation countermeasure Network (Cgan) is an extension of the original Gan, the generator and the discriminant are added additional information y is the condition, Y can make arbitrary information, such as class information, or other modal data. As shown in Figure 1, the conditional Gan is achieved by conveying the additional information y to the discriminant model and generating the model as part of the input layer. In the generation model, the priori input noise P (z) and the conditional information Y are combined to form the joint hidden layer characterization. The antagonism training framework is quite flexible in the composition of the implicit layer representation. Similarly, the objective function of conditional Gan is a two-person minimax game with conditional probabilities (Two-player minimax game):

network structure of Cgan
3.3 Experiments 3.3.1 Mnist Data Set Experiment

In Mnist, the condition of the class label (One-hot coding) of the training condition Gan, according to the label condition information, the corresponding number can be generated. The input to the generated model is a 100-dimensional, uniformly distributed noise vector, and the condition variable y is the one hot code for the category label. Noise z and label y are mapped to hidden layers (200 and 1000 cells), and all cells are joined before mapping to the second layer. Finally there is an output of the sigmoid generation model (784-D), the single-channel image of the 28*28.
The input of the discriminant model is the 784-D image data and the condition variable y (one hot encoding of the category label), and the output is the probability that the sample comes from the training set.
3.3.2 Multi-modal learning for automatic image labeling

Automatically label images: Automated tagging of images, using multi-label predictions. The distribution of tag-vector in the image feature condition is generated using the condition Gan. Datasets: MIR Flickr 25,000 DataSet, language model: Train a Skip-gram model with a 200-d word vector.
"Generate model input/output"
Noise Data 100 D =>500 dimension
Image features 4096-dimensional =>2000 dimension
These units are all combined to map to a 200-dimensional linear layer,
Output generated word vectors (200-D word vectors)
"discriminant model input/output"
Input:
500-dimensional word vector;
1200-D image features
??? The condition of the generative and discriminant input y, the dimension is not the same??? One is a 4096-dimensional image feature, the other is the? Vector _???
As shown in Figure 2, the first column is the original image, the second column is the user labeled tags, the third column is generated by the model G tags.
3.4 Future Works

1. Propose more complex methods to explore the details of Cgan and analyze their performance and characteristics in detail.
2. Each tag currently generated is independent of each other and does not reflect richer information.
3. Another remaining direction is to construct a joint training scheduling method to the school language model Reference

[1] Mirza M, Osindero S. Conditional generative adversarial nets[j]. Computer Science, 2014:2672-2680.
[2] Goodfellow, I, Mirza, M., Courville, A., and Bengio, Y. (2013a). Multi-prediction deep Boltzmann machines. In advances in neural information processing Systems, pages 548–556.
[3] Denton E L, Chintala S, Fergus R. Deep generative Image Models using a Laplacian Pyramid of adversarial networks[ C]//advances in neural information processing systems. 2015:1486-1494.
[4] Radford A, Metz L, Chintala S. Unsupervised representation learning with deep convolutional generative Adversaria L NETWORKS[J]. ARXIV preprint arxiv:1511.06434, 2015.
[5] Goodfellow Ian, Pouget-abadie J, Mirza M, et al. generative adversarial nets[c]//advances in neural information P Rocessing Systems. 2014:2672-2680.