I am working on regenerating matrices. Those matrices represent 6 by 6 grid layouts. They contain values that show how the grid can be divided.
So , I am using disentangled variational autoencoders for such task. However , the results that I am generating doesn't look like the input or it is so far from the input.
According to what I have read on the internet, in order to have good results you have to increase the beta value more than 1 but in my case I got better results when I decrease beta.
I would like to know why I am getting such results.
This github link https://github.com/emnajaoua/beta_variational_autoencoders/blob/master/disentangled_vae%20(1).ipynb
contains the code that I am using.
Thank you in advance.
As far as i know, it depends on various things such as number of training samples, number of layer in your network, loss function, and one of the important factor is sparsity proportion and sparsity regularization.
I will recommend to use different loss functions and update those parameters, i hope you will achieve good results if your training data is precise and not noisy.
https://www.researchgate.net/topic/Deep-Learning
https://arxiv.org/pdf/1802.05814
https://arxiv.org/pdf/1702.02526
https://www.jeremyjordan.me/variational-autoencoders/
https://www.biorxiv.org/content/biorxiv/early/2017/11/05/214247.full.pdf
https://www.biorxiv.org/content/biorxiv/early/2018/03/08/278739.full.pdf
As far as i know, it depends on various things such as number of training samples, number of layer in your network, loss function, and one of the important factor is sparsity proportion and sparsity regularization.
I will recommend to use different loss functions and update those parameters, i hope you will achieve good results if your training data is precise and not noisy.
Mohamed AbdElAziz Khamis , thank you for the suggested papers but I didn't find in them how to reparamterize my model that uses the B-VAE in order to have better results.
Amin Ullah , Thank you for the suggestions. Actually, I tried increasing the number of layers in the network and it has effected a bit the results positively.
I checked also the sparsity proportion in my data, I found out that my data is almost considered as dense and non sparse.
Concerning the preciseness and non-noisiness of the training data, I am randomly generating different 6 by 6 matrices for training data, So if you can have a look at the github code you will find the part of generating the data. You can check or recommend for me how to check if my data is precise enough.
https://stats.stackexchange.com/questions/183653/getting-the-right-starting-values-for-an-nls-model-in-r
https://www.uv.es/uriel/4%20Hypothesis%20testing%20in%20the%20multiple%20regression%20model.pdf
Article Data-driven spatial b value estimation with applications to ...
https://www.hindawi.com/journals/mpe/2012/640472/
Article Reparametrization-based estimation of genetic parameters in ...
In case of random generation means that you wouldn't have similar samples which can not be precise enough for generation new samples. so in this case after data generation you could try data augmentation techniques which will increase the number of sample as well as intra-similarity between samples. You can try the following data augmentation methods.
https://github.com/aleju/imgaug
https://github.com/mdbloice/Augmentor
Amin Ullah Thank you for your suggestion. I will try that data augmentation then. It makes sense in my case.
- Badges
- Science topic