Questions tagged [generative-adversarial-networks]

For questions related to generative adversarial networks (GANs), introduced in the paper Generative Adversarial Nets (2014) by J. Goodfellow et al. A GAN is composed of a discriminative model (D) and a generative model (G). The discriminator D needs to distinguish between data generated by the generator G and data in the training set, while the generator G needs to generate data such that the discriminator D is not able to accomplish its task.

Filter by
Sorted by
Tagged with
0 votes
0 answers
14 views

Idea for generating time series with irregular time-intervals with GANs

I want to model time-series with irregular time-intervals using GANs. Think of the following (short) data sample ...
-1 votes
1 answer
27 views

Combination of VAE with GAN

I am going to implement a lecture which it aims to generate new images. It uses a variational autoencoder to produce latent vector and then feed it to a gan network as input. My question is, in ...
0 votes
0 answers
15 views

What part of the WGAN-GP loss should consist of the gradient penalty?

I'm trying to finetune a WGAN-GP-based model. The discriminator loss is logically partially made up of a gradient penalty. However, in much of my testing, the gradient penalty makes up a large part of ...
1 vote
0 answers
38 views

Is it possible to combine DDPM with GAN?

From what I understand in GAN, the main idea is that you have a generator and a discriminator network that are "competing" with each other. The generator trying to make images that the ...
0 votes
1 answer
35 views

Using GANs to generate data augmentations for YOLOv5

I was building a YOLOv5 object detection model, and was looking into researching synthetic methods like GANs to increase the size of my training set in an unsupervised manner. I know that few-shot ...
  • 3
0 votes
1 answer
24 views

Can any GAN's utilize labels in their datasets while they are training?

It seems to me that the Generative Adversarial Networks have a practical issue when trying to reproduce some of their output images For example, as you can see https://www.youtube.com/watch?v=...
0 votes
0 answers
6 views

How does the 'moment's matching trick' in GAN training improve the diversity of the generated samples?

I was investigating the TimeGAN code, when i stumbled across the 'moments loss' component. In one of the issues, the author states that this is a 'moment's matching trick' used 'to improve the ...
0 votes
1 answer
32 views

Best way to generate a human face over a face generated by FaceFormer framework?

FaceFormer framework generates a talking face from audio, focusing on the lip and face movement when a person talks. Now from that what would be the best way to generate a human face on top of that? I ...
1 vote
2 answers
62 views

GAN with multiple discriminators

I am looking for literature recommendations regarding GANs with multiple discriminators. In particular, I am looking for examples where each discriminator has a slightly different learning objective, ...
0 votes
0 answers
19 views

Why would one still use a traditional GAN architecture or WGAN architecture instead of a WGAN-GP architecture?

I've been diving into the literature of GANs, and quite early on, I was pretty convinced that WGAN-GPs were the way to go. The WGAN-GP architecture is, as far as I know, theoretically and empirically ...
0 votes
0 answers
13 views

Adding MNIST images by using them as channel inputs

I'm trying to create a generative neural network that can offer "basic sum" mathematical solutions using the MNIST dataset from a conditional input. I've curated a dataset of MNIST examples ...
  • 1
0 votes
1 answer
149 views

Why don't OpenAI train a deep learning model to identify correct and incorrect information in ChatGPT's responses?

I'll preface this by saying that I have little experience in artificial intelligence, so this might be a naive question. However, in light of the recent controversy surrounding ChatGPT's inability to ...
  • 109
0 votes
1 answer
45 views

Why does data augmentation using synthetic data generated by one model improve the performance of another model?

I understand from articles like this one that synthetic data generated by one model based on real data can improve the performance of a second model. Can anyone help me understand the intuition behind ...
-1 votes
1 answer
28 views

Generating synthetic time series data with limited data

I would like some opinions on my current situation. I have a set of time series data that I want to forecast. The data however is not very long (around 500 rows) so I was looking into generating many ...
1 vote
1 answer
70 views

Can you extrapolate outside the latent distribution for GANs?

I was wondering what happens when you extrapolate out of the latent space distribution (noise vector) for a Generative adversarial network (GAN). Can anybody explain this?
  • 35
0 votes
2 answers
97 views

Is Relativistic GAN better than WGAN-GP?

I am currently reading the ESRGAN paper and I noticed that they have used Relativistic GAN for training discriminator. So, is it because Relativistic GAN leads to better results than WGAN-GP?
0 votes
1 answer
58 views

What does the adversarial loss in a GAN represent?

I'm working on Pix2Pix an image-to-image translation GAN, and I noticed that there is an adversarial loss implemented using BCE, and a L1 loss implemented using MAE. I know L1 loss represents the ...
0 votes
0 answers
21 views

Why did the authors of D2GAN propose two discriminators using KL and reverse KL divergence losses instead of one discriminator using JS divergence

I have stumbled upon the D2GAN paper as part of my research and I am finding myself extremely confused by the fact that instead of using the JS divergence to capture both KL and its reverse, they ...
  • 101
2 votes
0 answers
287 views

What is the difference between a diffusion model and GANs?

Recently, I hear a lot of people claiming that diffusion models beat GANs, also providing less training time. I've searched about these two type of models, and I am confused, because somehow they both ...
0 votes
0 answers
26 views

Why GANs trained on QM9 dataset produces drug-like molecules?

Why the GANs (Generative adversarial Network) trained on QM9 dataset (contains 134K molecules but none of them is complete to be eligible for drug-like molecule) produces drug-like molecules. sine ...
0 votes
0 answers
15 views

Generator for a string containing letters and digits with unknown output

in preparation for a new project i would like to ask for some help finding an approach. The Task: Some kind of material-number in System A (numbers&letters of length X) need to be mapped to ...
  • 1
2 votes
1 answer
114 views

Clarification on GANs for text generation

A GAN-like architecture for text generation is proposed in 'Generative Adversarial Networks for Text Generation'. The setup is the following: The generator of the GAN is proposed to be a recurrent ...
0 votes
1 answer
46 views

Does the Weights of Discriminator get updated when traning Generators in GANs?

When we train the GAN we usually train the discriminator first then the generator, first we stop the generator from updating its weight by removing it from the computation graph, using fake_image....
0 votes
1 answer
40 views

Is the discriminator of a GAN network embedded in VAE?

From what I understand, a Generative Adversarial Network (GAN) is composed of an encoder (generator), some synthetic data (fake data) and a discriminator that will penalize any distinguishable real ...
  • 3
0 votes
1 answer
35 views

U-Net Maxpooling vs Convolution

Hello I'm implementing a CycleGAN and most of the other implementations I've seen on the internet use Convolution with stride 2 instead of a Maxpoolinglayer for downsample. On to my question, why ...
  • 1
0 votes
1 answer
53 views

Parametric vs Non-parametric generative models

I have a little perplexity trying to distinguish parametric vs non-parametric generative model. In my understanding, a parametric generative model would try to learn the probability density function ...
0 votes
0 answers
111 views

Why Pix2Pix loss function does not lead to overfitting?

In the Pix2Pix paper, the loss function is described as $\ \mathcal{L}_{cGAN}(G,D) + \lambda \mathcal{L}_{L1}(G)$ . Where the L1 loss in the model is the difference of pixels between the generated ...
0 votes
0 answers
21 views

What is the logic behind using a trained classifier's gradients to synthesize controllable image?

In the controllable image synthesis, we are manipulating a noise vector z such that our generator ( in our GAN model ) creates images that the desired feature exists. For instance, take the feature of ...
2 votes
1 answer
291 views

What is being optimized with WGAN loss? Is the generator maximizing or minimizing the critic value?

I am kind of new to the field of GANs and decided to develop a WGAN. All of the information online seems to be kind of contradicting itself. The more I read, the more I become confused, so I'm hoping ...
1 vote
0 answers
41 views

Is there a name for this model?

I have an image autoencoder model trained as follows: Step 1) train a GAN to obtain a generator capable of drawing from the data manifold by sampling a normal distribution in latent space Step 2) ...
0 votes
1 answer
42 views

What is the meaning of $p_{\text {data }}(y)$ in the CycleGAN?

In the original CycleGAN paper, on the second page, there is a sentence that I didn't quite understand In theory, this objective can induce an output distribution over $\hat{y}$ that matches the ...
1 vote
2 answers
138 views

What is the domain of the discriminator of a GAN?

I've read that the discriminator $D$ validates an image $D(x)$, where $x$ is either a real image or a fake one created by the generator, i.e. $ D(G(x))$. What does the function of the discriminator ...
0 votes
1 answer
556 views

Why do we train the discriminators k times but train the generator only 1 time in a iteration in GAN?

In this paper https://arxiv.org/abs/1406.2661 , the codes for training a gan are : Why do we train the discriminator for $k$ steps while the generator only for $1$ step? Why not the other way around?
  • 3
0 votes
0 answers
23 views

Why even-sized kernels are used in upscaling layers?

I have noticed that UNet and many GANs uses even-sized kernels in the upscale part of the model. I have read that at least in the GAN situation one of the reasons why we use even-sized kernels is that ...
  • 133
2 votes
1 answer
59 views

Why do we use the same parameters for the joint, marginal and conditional distributions in VAEs?

I've noticed in several resources on variational autoencoders (for example the wikipedia article), we use the same parameters theta for the prior, likelihood, posterior, etc distributions. For example ...
  • 21
0 votes
0 answers
37 views

What are the benefits of using multiple convolutions, as opposed to one, before the pooling layer in a U-Net?

I have seen U-Nets that use a single convolution before the pooling operator and some that use two or more. My question is, what is better? Or what are the benefits of using more or less convolutions?
0 votes
1 answer
57 views

Is the "Helvetica scenario" mentioned here related to Artificial Intelligence?

Consider the following sentence from the original GAN paper titled Generative Adversarial Nets in particular, $G$ must not be trained too much without updating $D$, in order to avoid "the ...
  • 3,451
0 votes
0 answers
28 views

Why GAN is not as useful in tabular settings

This is inspired by a previous question Why are GAN models not heavily used for NLP? One answer says that the GAN is not as useful outside the image domain. Recently my friend and I are trying it out ...
0 votes
1 answer
86 views

Training a GAN after after evaluation metric reaches minimum

I am training a StyleGAN-3 using one of the pre-trained models. At some point, roughly halfway through the 5000 kimg recommended for fine-tuning, the FID50K score starts oscillating around a minimum ...
  • 3
1 vote
0 answers
41 views

Does all GAN's in literature need to satisfy the properties of objective function of initial GAN? [closed]

Consider the following value function of the initial GAN $V(D, G) = \mathbb{E}_{x \sim p_{data(x)}} [\log D(x)] + \mathbb{E}_{z \sim p_z(z)} [1- \log D(G(z))]$ The min-max game on the value function: $...
  • 3,451
3 votes
2 answers
68 views

Why doesn't the inception score measure intra-class diversity

It's mentioned here that there is no measure of intra-class diversity with the inception score: If your generator generates only one image per classifier image class, repeating each image many times, ...
0 votes
0 answers
34 views

Finer control over the distribution of output from a GAN

I have cast some data into an image (25 x 56) to work better with existing tools, and then used a CNN to train a GAN using the Wasserstein loss with gradient penalty to generate new samples of the ...
  • 1
0 votes
0 answers
16 views

How is state size obtained/calculated?

I am looking at the Generator portion of the Pytorch tutorial for Generative Adversarial Networks and I am confused as to how the last 2 dimensions are obtained for what is ...
1 vote
1 answer
176 views

What are the steps to derive the original GAN loss function from the generalized version?

I am trying to understand how the loss function from the original GAN paper $$\min_{G} \max_{D} V(D, G)=\mathbb{E}_{\boldsymbol{x} \sim p_{\text {data }}(\boldsymbol{x})}[\log D(\boldsymbol{x})]+\...
  • 13
0 votes
0 answers
241 views

How exactly is modulation and demodulation layer in StyleGAN2 implemented?

So from the paper Analyzing and Improving the Image Quality of StyleGAN We know that the naive way to implement the stylegan2 Conv2DMod is to compute the Style vector which has the dimension of ...
2 votes
1 answer
298 views

Does a colour consistency loss in neural networks (cycleGAN) make sense?

My neural network takes an image as an input and outputs another image. It's the generator of a cycleGAN. I would like to add (to the discriminator loss, the ...
0 votes
1 answer
181 views

GANs inputs normalized and generator only outputs in [-1; 1]

I'm currently coding a GAN on the dataset MNIST. I'm using the following code to transform my data: ...
0 votes
0 answers
94 views

Interpretation of the loss function of Wasserstein GAN: the lower the better?

After following this interesting collection of tutorials for GANs https://www.youtube.com/playlist?list=PLhhyoLH6IjfwIp8bZnzX8QR30TRcHO8Va I've been playing around experimenting Wasserstein GAN with ...
4 votes
1 answer
2k views

Why are GAN models not heavily used for NLP?

I am wondering why there has not been more usage of GANs for NLP. I know there has been research on the subject (The Google Scholar page for the subject is here). Are there any specific reasons why ...
2 votes
1 answer
225 views

Why do we use a linear interpolation of fake and real data to penalize the gradient of discriminator in WGAN-GP

I'm trying to better frame/summarize the formulations and motivations behind Wasserstein GAN with gradient penalty, based on my understanding. For the basic GAN we are trying to optimize the following ...

1
2 3 4 5