Questions tagged [generative-models]
For questions about models designed for generating new data (or generating samples from a probability distribution).
168
questions
0
votes
1
answer
43
views
Purely extractive Language Model
Given an email thread, I am trying to extract the body of the most recent email.
I used to do that with rules. Now I am testing Large Language Models (LLM) to see if I they provide a less ad hoc ...
0
votes
0
answers
7
views
Assessing performance of a generative model (images) using Jensen-Shannon Divergence
Background
I am currently working on comparing two high-dimensional image datasets (MNIST) of shape $\sim(50000, 256)$ to evaluate the performance of a generative model, and have attempted to use ...
1
vote
1
answer
46
views
Unsupervised Machine Translation System Using Variational Autoencoder Models
I want to work on an unsupervised machine translation system using a variational autoencoder. I did a literature review but didn't find any related work, and most of the work is based on denoising ...
0
votes
0
answers
18
views
Improving Normalizing Flows Accuracy
What are some techniques one might use to improve the accuracy of normalizing flows? I am training a flow in a high-dimensional space but it seems like there's always at least one or two dimensions in ...
0
votes
0
answers
19
views
In the GAN objective function, why do we first do we first find the D(x) that maximizes the objective function and then maximise wrt the generator?
The GAN objective function is optimised like this: argmin(argmax(L(G,D))) where the argmax finds the D (Discriminator) that maximises L(G,D). Why is it not the other way around, i.e. argmax(argmin(L(G,...
1
vote
1
answer
213
views
issue loading the ckpt file PytorchStreamReader failed reading zip archive: failed finding central directory
I am trying to load the ckpt file and getting error
PytorchStreamReader failed reading zip archive: failed finding central directory
Here is the code
...
2
votes
3
answers
430
views
torch cuda not able to identify gpu on aws g4dn.xlarge
I have created an EC2 instance with GPU g4dn.xlarge and launched it. I want to run some ...
1
vote
1
answer
57
views
Why is T5 often used in text-to-data for text prompt encoders?
In the text-to-data(music, image, audio, etc.) generative AI field, one method of encoding text prompts is to use pre-trained language models. Such an approach was used in research on Moûsai [1] and ...
0
votes
0
answers
11
views
Generating synthetic labeled data (sampling from p(x,y))
I'm working on a toy problem. Consider a dataset that consists of 1-D vectors (waveforms) that contain noise, except for one prominent spike. Denote the waveform by $\vec{x}$, and let the coordinate ...
1
vote
0
answers
16
views
Avoid overfitting to noise by a noise penalty approach instead of early stopping?
I came across this article on deep learning for computational MRI and found an interesting sentence "However, early stopping has to be performed to not overfit to the noisy measurements." ...
0
votes
1
answer
55
views
What Deep Learning model to use in this spectroscopy task?
I have a task to be solved. There are energy measurements over the square area 40x40. One measurement consists of values : x, y and the energy. The all area is almost whole covered with data (a few ...
-1
votes
1
answer
18
views
Build a topic model without data?
I need to come up with a topic model, without any labelled dataset, the model should also be multilingual, thinking of using LLM's as they are accurate and awesome but if Im to build one on my own how ...
0
votes
0
answers
32
views
Is it a good idea to use attention in VAEs for image generation?
There are research papers and codebases on GitHub that deal with VAEs for image generation on popular datasets like CelebA, etc. While surfing through Google Scholar I found self-attention and other ...
0
votes
0
answers
16
views
How to train a VQ Gan Based on Pretrained Language and Image Encoder?
I have trained the dual encoder model (separate language and vision encoders) using the ideas here.
However, this looks like an image look up model rather than an image generation model. What would ...
0
votes
0
answers
11
views
Any ideas on how to generate code based NLU Data
I am working on a Generative AI project. I wish to explore stuff possible using "Code inputs".
...
0
votes
0
answers
38
views
How to train a custom sentence completion model using tensorflow?
What I have
A small corpus of English sentences (about 60,000).
The Task
Sentence completion model. Basically, something like
Input: "Paris is the"
Output: "Paris is the capital of ...
0
votes
0
answers
414
views
How to Use Generative AI for Time Series Forecasting?
What I have
A time series dataset of time stamps (hourly resolution), some covariates (like store foot-traffic) and items sold.
What to forecast
Number of items sold for next 24 hours, i.e. 24 ...
2
votes
2
answers
94
views
Edit friendly DDPM noise space
I was reading this paper, "An Edit Friendly DDPM Noise Space: Inversion and Manipulations". In page no. 4, they have mentioned that in DDPM, noise maps of consecutive steps are highly ...
0
votes
0
answers
22
views
How does the use of 1x1 convolutional layers represent permutation in the GLOW model?
I am currently reading the GLOW paper (found here) and I can not understand how the authors claim that the use of 1x1 convolutional layers is equivalent to permutation holds true.
I understand how a ...
0
votes
0
answers
26
views
Help with running topology GAN (TopoGAN)
Link: https://github.com/TopoXLab/TopoGAN-ECCV2020
Sorry if this is the wrong place to ask, but I've been looking for help on how to work this out for a long time as I don't have much experience in ...
0
votes
0
answers
19
views
CGAN - Odd Distribution gap, failure of convergence?
I am trying to train on some 1 dimensional data (675 samples, its very expensive to get more) and trying to match the distribution seen here:
There are labels from 1-3 as to associated with the noise....
0
votes
0
answers
16
views
Seeking Solutions for Generating Text Descriptions from Diagrams, Infographics, Charts, etc
I am currently on a quest to find an efficient way to generate meaningful text descriptions (or alt text) from visual representations such as diagrams, infographics, charts, plots and the like. ...
0
votes
0
answers
29
views
Generator loss keeps increasing while discriminator keeps decreasing
I am trying to build a GAN to generate LEGO images however my generator is not working at all. I have tried changing the learning rates but it caused the loss to go even more higher, sometimes into ...
1
vote
0
answers
137
views
Common sense fixes to a buggy diffusion model that won’t overfit one sample?
hope this question is in the right place. I’m working with a toy diffusion model to generate points e.g learning a Swiss roll which to me is a basic use case that I wanted to start with.
My model is ...
1
vote
0
answers
20
views
Which approach to determine the size of an object to be placed given the sizes of the existing objects in a scene?
I am working on an automated approach to object-based data augmentation. The goal of the approach would be to add a selected object to an existing image. To automate this task, information is needed ...
0
votes
0
answers
21
views
How to refine sampling of points via averaging?
I am working with a generative model which is generating points that are less accurate than I would like. I have strong reason to believe the errors should average out along a particular axis (to give ...
1
vote
0
answers
18
views
Generating artificial training data with encoder and classical algorithm
I would like to know if this idea has been tried before, and if so, where I can find more information about it.
This is an approach to generating artificial training data for segmentation tasks using ...
1
vote
2
answers
99
views
Why is cycle consistency loss alone not sufficient to produce meaningful output?
Imagine an adaptation of CycleGAN, in which the discriminators were removed in lieu of using only cycle consistency loss. Well, it turns out that the original authors of Cycle Consistent Adversarial ...
0
votes
0
answers
11
views
Why does this cGAN model perform poorly when trained on a different machine?
I was sent a cGAN model python file from a friend + the dataset he used to train this model. For him, the model trained succesfully & was able to generate very accurate images. These were his ...
0
votes
0
answers
12
views
How to generate synthetic feature instead of synthetic image using Diffusion-based generative Model?
Is it possible to pass some extracted features from the pre-trained ResNet model to a diffusion model for training and further generate synthetic features instead of images like GAN or VAE?
P.S. I ...
0
votes
0
answers
23
views
Is a conv transpose layer equivalent to a padding layer and regular conv layer
Is a 2d convolution transpose layer equivalent to a upsampling layer that inserts 0s between rows and columns, then a regular 2d convolution layer? If so, why is it usually not implemented as such (i....
1
vote
0
answers
52
views
NANs, Infinities, and very large losses with normalizing flows
I am new to normalizing flows and have been trying to use them with a high-dimensional dataset, and I have been running into very large numbers and errors with sampling that don't occur when I use a ...
0
votes
0
answers
38
views
Is this overfitting? (generative model)
I am working with a generative method, and the network seems to perform well on training data and slightly less well on test data, but the generated data is somehow significantly worse than either of ...
0
votes
1
answer
288
views
Transpose Convolution Output Size
I have been learning GAN (Generative Adversarial Networks) lately and having a hard time understanding the output size for transpose convolution. Let's say I am using a Tensor of [1, 64, 1, 1] as an ...
0
votes
0
answers
8
views
What models/techniques can I use to generalize industry specific datasets?
I have a few dictionaries pertaining to different industries (ie. tech, manufacturing, education, etc.). These dictionaries map phrases and keywords to a sentiment score. I'd like to create a ...
0
votes
0
answers
9
views
Adaptive Generation in EBMs
I have a question about one Wikipedia article about EBMs: Why does it adapt without training?
EBM generators are implicitly defined by the probability distribution, and automatically adapt as the ...
0
votes
1
answer
73
views
Best practices for serving user-specific large models in a web application?
First execuse any naive statement you may find below, i'm a newcomer to the field.
How do web applications that integrate fine-tuning of large machine learning/deep learning models handle the storage ...
0
votes
0
answers
12
views
Normalizing flows with truth values?
I am interested in using normalizing flows to map to a certain distribution of points, but I want to make sure that the distribution is not just of the correct general shape, but also that specific ...
0
votes
0
answers
19
views
How to deal with degeneracies in data
I am working with a dataset where I want to predict some features (let's call them G) from some other features (we'll call them W). The problem is that there is no one-to-one mapping from W to G. In ...
1
vote
1
answer
119
views
Autoencoder: How should hidden layer be used?
I'm building a variational autoencoder to generate faces. I'm using gray-scale images with the size 30x30. I started with a very simple model:
Input Layer, 900 nodes, values 0-1
Latent Space, 10 nodes ...
0
votes
0
answers
56
views
Sigmoid Activation Function (Output layer) Alternative
I have a Convolutional-VAE architecture where the target images are in the range [0, 1], their pixel values. To synthesize/reconstruct images in this scale, I am using a sigmoid activation function in ...
0
votes
0
answers
63
views
Practical example of difference between p(y|x) and p(x|y)
I've been reading about the difference between generative models and discriminative models. I know that for generative models we learn the joint probability p(x,y) or just p(x|y) and p(y). For a new ...
3
votes
1
answer
102
views
Why is it an advantage "that Markov chains are never needed" to obtain gradients?
In the original GAN (Generative Adversarial Network) paper, Generative adversarial networks by I. Goodfellow, J. Pouget-Abadie, M. Mirza et. al. they state an advantage of the GAN is "that Markov ...
0
votes
1
answer
34
views
How to generate structured parameters from a spectrogram?
Say I have an algorithm that accepts as input structured parameters of the following format, generates an audio clip and then a 512x512 spectrogram out of it:
...
1
vote
0
answers
96
views
How to implement simple VAE with sparse tensor in Tensorflow
thanks for reading.
I have been attempting to train a simple VAE on very sparse 2D and 3D data. So far I have been training using dense tensors which - I think - is resulting in horrible training due ...
1
vote
1
answer
692
views
How to select relevant columns from a dataset with many features
I have a dataset with a large number of potential features (>100) and I am interested in finding a relatively small subset of these (maybe on the order of 5, or 20) features which is best suited to ...
0
votes
1
answer
1k
views
KL divergence loss first decreases and then increases in VAE training
I am training a VAE on CelebA HQ (resized to 256x256). The training is going well, the reconstruction loss is decreasing and reconstructions are also meaningful. But, the problem is with KL divergence ...
1
vote
0
answers
30
views
ML techniques for mathematical inverse approximations
I have some inputs and outputs of a set of functions, and I want to be able to find/approximate any given input vector from its corresponding output vector (In other words learn the inverses of these ...
1
vote
0
answers
15
views
In which way GAN generator transforms the data(for transforming a noise to the data)?
I have the problem: I understood how GAN works in general, but I need information how it work detailed. The part I don't understand is how the random noise at input is transformed to data on the ...
0
votes
1
answer
236
views
Are genetic algorithms considered to be generative models?
My understanding is that these sorts of algorithms can evolve/mutate data to hone in on specific desirable areas in large/difficult to search parameter spaces. Assuming one does this successfully, how ...