2

Most of the tutorials online seem to use VAEs to generate images and use CNNs to generate data. I am working on a game with multivariate data consisting of character position and the character attributes to train a VAE.

I was wondering if it was possible because I have tried training the network with pytorch but I continue to get poor results with KL Divergence term decreasing to 0.

Any help would be really appreciated.

Alex
  • 21
  • 1
  • it probably has nothing to do with pytorch and more likely to do with your data quality. but without any information on what your data looks like, it's impossible to tell. – Victor Ng Mar 03 '20 at 21:45
  • @VictorNg Could you elaborate more on data quality? What would be an example of a good data in this case? Sorry that I cannot share more information on how the data looks like due to regulations. – Alex Mar 03 '20 at 22:26
  • well you can definitely use autoencoders to generate multivariate data. i was using it for fraud detection personally, but by drawing from the encoded distribution, you can generate new data. by data quality, I'm just saying it's difficult to help diagnose where the exact issue is if all we know is character position and character attributes and nothing about what the data looks like. – Victor Ng Mar 04 '20 at 02:44

2 Answers2

1

The CTGAN paper (github) use a VAE to generate multivariate synthetic data from tabular data. See section 4.5.

al0
  • 141
  • 5
0

Sure can look at this multivariate timeseries for example: enter image description here

It has nothing to do with pytorch, input is input, regardless. See this paper

Noah Weber
  • 5,609
  • 1
  • 11
  • 26