VariationalAutoEncoderExample problem

[Boeckie]

Hi,

I'm trying to reproduce the VariationalAutoEncoderExample results as displayed in the "Deep Learning" book on page 225 and page 227.

I've experimented with the parameters as defined in the source, as defined in the book and as defined on http://blog.fastforwardlabs.com/2016/08/12/introducing-variational-autoencoders-in-prose-and.html.

For some reason, the different digits in my scatter-plot are not properly clustered (huge overlap). Did anyone succeed in reproducing the result from the book?

All feedback is appreciated!

Thanks! Bart

[AlexDBlack]

I assume you've seen this? https://github.com/deeplearning4j/dl4j-examples/blob/master/dl4j-examples/src/main/java/org/deeplearning4j/examples/unsupervised/variational/VariationalAutoEncoderExample.java

As for the hyperparameters - I'd have to see exactly what you are running before I can make any comments.

[Boeckie]

Hi Alex,

Thanks for the feedback!

The code you referenced is indeed one of the experiments I ran. So far, the scatter-plot is not properly clustered.

Thanks, Bart

[AlexDBlack]

Here's what I'm getting from the example (for the last training iteration, via the slider)... you are getting something different?

[RobAltena]

Closing on inactivity.

[sercharpak]

Hello !! I'm running the VAE Example (dl4j-examples/src/main/java/org/deeplearning4j/examples/unsupervised/variational/VariationalAutoEncoderExample.java) as it is in the github. It runs properly but then when the visualization windows pop-up, they only properly show the first iteration (the rest of iterations just appear blanck (only the grid appears for the latent space plot and the generated MINST digits are all black).

Is this a normal behaviour_ Is there something to modify to the original github code ?

Thank you so much. This is awesome!

(I'm running in Windows 10)