Tensorflow implementation of Adversarial Autoencoders

github.com

64 points·conan7882··10 comments

For this particular implementation, what are the advantages of using it for supervised or unsupervised learning? (in general?)

And if it's just being used for the MNIST dataset, is there a particular reason for using it in one or the other fashion?

From what I can gather the supervised approach allows you to only learn the representation for style rather than which digit it is. The only reason to use one over the other is to demonstrate that it works, I guess?

Question: what framework is used mostly by academic researchers these days for DL?

Tensorflow, although PyTorch is gaining popularity.

Ok. I guess in this case the authors used TF because three of the authors work at Google.

This code is really good.

And impossible to use thanks to no license information.

MIT License, it was added after your post https://github.com/conan7882/adversarial-autoencoders-tf/blo...

Though, in the finest academic tradition, once you try to actually run it, you'll find that it silently depends on a separate library written by the author, which you'll have to find yourself.

(In this case "tensorcv", which is in a separate repository: https://github.com/conan7882/DeepVision-tensorflow )

That aside, I agree that it's an easier read than most ML code.

I though I have removed all the dependency on 'tensorcv'. It turns out I forgot the dataflow part. Now it should be run without 'tensorcv'. Thanks for pointing out.

Reply on news.ycombinator.com