Stack Exchange Network

Stack Exchange network consists of 175 Q&A communities including Stack Overflow, the largest, most trusted online community for developers to learn, share their knowledge, and build their careers.

Visit Stack Exchange

Questions tagged [gan]

Generative Adversarial Networks (GANs) are neural networks that are trained in an adversarial manner to generate data mimicking some distribution.

1
vote
0answers
11 views

Usage of dropout in convolutional GANs with batch norm?

In DCGAN, dropout is not used in either generator or discriminator. When using batch norm, are the benefits of dropout generally so marginal that is is not used? If it is used, in what circumstances?...
1
vote
1answer
21 views

Is a GAN's discriminator loss expected to be twice the generator's?

If a GAN generator has the same (but reversed) hidden layer architecture as the discriminator, is a the discriminator's loss expected to be approximately double the generator's? In the examples I'm ...
1
vote
1answer
29 views

How can we relate the concepts of GAN/cGAN in SRGAN? Is SRGAN a Conditional GAN?

I have been reading and looking at implementations of the SRGAN, from "Photo-realistic Single Image Super Resolution with Generative Adversarial Networks" paper. One thing that I noticed is that the ...
0
votes
0answers
12 views

Architecture used by author in StackGAN

I was going through this paper stackedGAN I somehow understood how it is working. But I wanted to know it's architecture so that I can implement it myself. I went through the implementation code of ...
0
votes
0answers
11 views

Why does the DCGAN output degrade with an increase in the kernel size?

Thank you for the explanation on the kernel size. I have been experimenting with the sample Generative Adversarial Network (GAN) code from the book on Deep learning with Python by François Chollet, ...
2
votes
1answer
16 views

Why are Generative Adversarial Networks classed as unsupervised

The title of the question is basically all I'm asking, but I should explain why GANs don't seem to be unsupervised to me! Here's my understanding of unsupervised learning: Unsupervised learning is ...
1
vote
0answers
20 views

What method to use for conditional density querying

I have a dataset of 3d poses each represented by 40 points (all relative to the central point). So my data has dimensionality 120. What is needed is to learn how build realistic pose, when positions ...
0
votes
0answers
24 views

DCGAN generator accuracy doesnt improve for high-res images

I trained a DCGAN on MNIST and CelebA dataset with 28x28 image size. Both the models were able to train successfully. I used many tips from https://github.com/soumith/ganhacks to make both the G and D'...
0
votes
0answers
13 views

How do GANs stay in sync?

What further research has been done since the introduction of GANs on the problem of keeping the generator and discriminator in sync, i.e. so one does not overpower the other?
4
votes
1answer
28 views

Convergence to gradient in limit of variance

I came across this equation in the original GAN paper (pg 2 https://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf): $$\lim_{\sigma \rightarrow 0} \nabla_{\bf x} \mathbb{E}_{\epsilon \sim \...
1
vote
0answers
22 views

Image generation based on sketch

Are there any instances of image generation models, where an image (a very rough sketch) has been used as an input and was then augmented. For example: This could be a rough sketch, which is then ...
0
votes
1answer
34 views

Generative model to generate hidden activations coming from a previously trained hidden layer

I need to train a generative model to generate vectors which resemble the activations of a particular hidden layer of a neural network which has been previously trained. In particular, the hidden ...
0
votes
0answers
12 views

What are the constituents of “distributions” in GANs?

We have a distribution for the Generator and the Discriminator, and we minimize their divergence, but how do the inputs (say, images) constitute a probability distribution? Or is the distribution ...
0
votes
0answers
117 views

Which GAN is the best for data augmentation?

I have around 200000 images and I want to augment the data by generating more of them. Images do not have classes, because they are the same object and are used for the task of object detection. Can I ...
2
votes
1answer
35 views

Prerequisites for Wasserstein GAN/Autoencoder

Can someone who read WGAN/WAE papers and understood Wasserstein part, could you share how you prepared necessary Optimal Transport background? The mentioned papers seem little tough if you don't have ...
1
vote
1answer
77 views

Why are GANs so innovative?

I've been reading about the importance of Generative Adversarial Networks (GANs), and I would like to double check that I understood correctly why they are so relevant. Before GANs, what people did ...
1
vote
1answer
84 views

What is the difference between Generative Adversarial Networks (GAN) and Generative Antagonistic System (GAS)?

What is the difference between Generative Adversarial Networks (GAN) and Generative Antagonistic System (GAS) in the neural network?
0
votes
0answers
45 views

Categorical data & Gaussian latent variables

I am learning about imposing structure on the latent variables in autoencoders. In that context I have looked at variational autoencoders (VAEs) and adversarial autoencoders (AAEs). This paper ...
0
votes
0answers
34 views

GAN generated images all similar per epoch

I'm working on a GAN using cifar-10 images. After each epoch I create 10 new random z noise vectors, and use them to create 10 images using the generator. All of the 10 images look very similar, but ...
-1
votes
1answer
44 views

Machine learning books covering neural networks / cnn / GAN [duplicate]

I'm not an expert in machine learning. Is there any textbook (with a decent amount of mathematical rigor) that cover the subjects neural network / convolutional neural network / GAN network? I've the ...
0
votes
0answers
106 views

Generative Adversial Networks (GAN) - Dimension of the Latent Space

I am trying to synthesis medical images with GAN. The problem is that my generator loss is very bad behaved: I read that if latent space dimension is not enough for representation of the true ...
0
votes
1answer
25 views

volume generation with GAN

I'm not a GAN expert, but I have a problem and I would like to understand if GAN could help me in some way. Essentially my problem is to convert a 3D grayscaled volume in another 3D grayscaled volume, ...
0
votes
0answers
18 views

Data Synthesis based on Deep Learning

Is there any open source tool available to synthetically generate a new dataset with the same statistics than the original one? The objective is to create a new tabular dataset that is private but ...
1
vote
0answers
114 views

GAN and NN for sparse data

I have a set of images which represent some correlated sparse data $x_1,\ldots ,x_n$. there are a number of specific pixels in the images which might hold value or not (with some probability), while ...
2
votes
0answers
73 views

GANs for image translation

I am training a generative adversarial network to perform style transfer from two different image domains (source S and target T)...
1
vote
1answer
52 views

Where's wrong in my reasoning behind upper bound for reconstruction error?

In the paper Mutual Information Neural Estimation, the authors derive the reconstruction error in BiGAN as $$ \mathcal R=E_{x\sim q(x)}E_{z\sim q(z|x)}\left[-\log p(x|z)\right] $$ where $q(z|x)$ is ...
0
votes
0answers
206 views

GAN training: Both G and D has very low loss

I am training a GAN. At the beginning the generator has a very high loss, which converges over time. After some time, the image quality seems pretty good, but both the generator and discriminator have ...
2
votes
1answer
71 views

In Ian Goodfellow et al's paper “Generative Adversarial Networks”, why do they specify that they do not need a Markov chain or inference network?

In Ian Goodfellow et al's paper Generative Adversarial Networks, they state, "There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of ...
1
vote
1answer
164 views

When should I stop a training of WGAN model?

The loss function of the WGAN is a continuous one. It doesn't have a convergence point. I don't really understand when we should stop the training.
2
votes
0answers
27 views

Any Asynchronous Training Methods for GAN?

GAN sometimes get really unstable with the high dimensional data. Can we train GAN is Asynchronous manner? It's like we have one master Generator and Discriminator. But we actually update it ...
3
votes
1answer
311 views

Why don't GAN generators vastly overfit?

It seems that, if the GAN generator is simply mapping noise to a value which should be as indistinguishable as possible for the discriminator from the real data, the generator could simply learn to ...
0
votes
0answers
156 views

GAN training : spikes in loss?

I came across the following situation during the training of a DCGAN on quite a complicated dataset : In many other cases of complicated datasets, loss G would slowly grow until at some point failure ...
0
votes
0answers
47 views

What does the shape of the optimal discriminator mean in GAN?

I know why the optimal discriminator is $D^{*}_G(x) = \frac{p_\text{data}(x)}{p_\text{data}(x) + p_g(x)}$ in Generative Adversarial Networks, but don't know how to intuit the state when the ...
1
vote
1answer
666 views

Can a GAN be used for tabular/vector data augmentation?

Can a generative adversarial network (GAN) be used for data augmentation (i.e. to generate synthetic examples that are added to a dataset) for data that is tabular/vectorized (i.e. not an image)? Are ...
2
votes
0answers
622 views

Face morphing using GAN

I need to morph multiple faces. Here is the desired result: I already tried to find facial landmarks using dlib shape predictor and then morph multiple faces, but unfortunately sometimes shape ...
2
votes
1answer
168 views

Why we train the generative model “indirectly” in GAN(generative adversarial networks)?

In the simple GAN here, I noticed when we train the generator, we are not directly training it by mapping the noise input (length 100 vector) to an image (28*28 matrix). Instead, the author is using ...
0
votes
0answers
190 views

Conditional GANs - conditioning scheme in TFGAN

In the 2014 paper on conditional GANs, https://arxiv.org/abs/1411.1784 the latent space vector z is augmented by a conditional variable, e.g. in the simplest case just the class label. There are ...
0
votes
0answers
20 views

$z$ dimensional restrictions and thumb rules in GAN's generator

A couple of questions on the dimensions involved in GAN's Generator. I am looking at various different papers (such as this one) which describe the generator as "mapping from a low dimensional ...
1
vote
1answer
30 views

Normalisation of an 'image' when pixel intensities are unbounded

I have a dataset from particle physics of 3D 'images' containing three spacial coordinates and one energy coordinate. I want to train a generative model for generation of a similar dataset. I am using ...
1
vote
0answers
2k views

Generative Adversarial Networks (GAN) to generate time series data

I see that there are cases of GANs used with Time Series. The paper Recurrent (conditional) generative adversarial networks for generating real-valued time series data says that they generated ...
0
votes
0answers
102 views

statistical testing - validate generative adversarial network results

I am working with clinical data (measurements) of patients. The patients are divided into two groups. One group has a certain illness and the other group doesn't. I now have trained a GAN with the ...
3
votes
1answer
205 views

how to stabalize GAN learning

I implemented a generative adversarial network. The generator and the discriminator both seem to learn independently. I put them together in the following manner : pre-train the discriminator to ...
0
votes
0answers
29 views

Confused on all the Bidirectional convolutional LSTMs variations, need intuition on picking the correct one

In my current personal project, I am trying to mimic a paper to generate moving MNIST dataset frames using CLSTMs. In the paper https://arxiv.org/pdf/1611.09904.pdf, they used standard bidirectional ...
0
votes
0answers
140 views

proof of the objective for generative adversarial network

There is any explanation or proof why the objection function of GANs yields to desired result? I have been reading some papers and tutorials regarding GANs and its variations, and I still have a ...
0
votes
3answers
479 views

What is the precise definition of unsupervised learning?

Let's look at a special case: Generative Adversarial Networks (GANs). (For those who don't know what a GAN is: for this purpose they are two neural networks that are trained using user generated ...
0
votes
0answers
392 views

CYCLE GAN : should we train Discriminator or Generator first?

As described content from this blog, when we train GAN, we should start with train Discriminator first, then freeze Discriminator and train Generator for each iteration. But when i tried to implement ...
1
vote
0answers
57 views

How to insert feature vectors as additional channels in conditional DCGANs

I understand that in a fully-connected GAN you can simply concatenate the flattened image and feature vector as input for the network. For convolutional GANs I've read that you should add the feature ...
2
votes
0answers
103 views

Generative Adversarial Networks - Gradient saturation

This is the value function from the GANs paper: The authors explain that this equation "may not provide sufficient gradient for $G$ to learn well", because early in the learning process the ...
0
votes
0answers
384 views

Transfer learning and Generative adversarial networks

I need to do image classification. I have small data, so I need to use Transfer learning. And also, I have some negative samples on my data, so I will use Gans. How can I combine GANs and transfer ...
1
vote
0answers
242 views

Reverse GAN, changing the distribution of the data to normal noise

I have a dataset which contains the data from 10 classes. The classes look well-separated that the accuracy rate of different classifiers is more than 95%. The goal is to change the data distribution ...