17:53 < ShikharJ> zoq: I wanted to discuss some ideas I thought might be of interest regarding the Deep Learning modules.

18:00 < zoq> ShikharJ: Great, please go ahead.

18:04 < ShikharJ> Talking in respect to GANs, Kris had suggested the development of the Wasserstein GAN (which I think is the latest one). Do you see a requirement for mlpack for this, or would you rather like to see other variants like DCGAN or SeqGAN?

18:08 < zoq> DCGAN is definitely an option, at the end some of this models share common structures, so if I like to implement Wasserstein GAN it would be much easier for me if I could reuse parts from an existing model e.g. DCGAN. So if you like to go for another model that'S absolutely fine.

18:13 < ShikharJ> Ah, right. Again building up on his suggestions, he had mentioned that his task for the previous summers was working up on StackGAN as well, which I guess wasn't completed? He had later suggested the implementation of a stacking module for RBMs and GANs, which I couldn't comprehend. Could you shed some light on that?

18:19 < zoq> We didn't get to the StackGAN part, since we encountered stability issues in the training step, so there is no existing code. I'm not sure about the last part, I guess the intention was to use the RBM for the energy function, which might be a neat idea.

18:25 < ShikharJ> Ah, I'll reach out to him for confirmation. How many modules do you think would be achievable in a GSoC length of time of 3-4 months? If I'm solely focussing on GANs, I'd like to work on pretty much all of the ones I mentioned, but then I'm not sure if these three (and maybe the stackGAN as well) would be over-ambitious or not.

18:28 < zoq> Yeah, I think all three might be too ambitious, a good idea would be to propose to e.g. implement two or one over the summer and to work on another one if there is time left. Note, writing good tests is something that takes a lot of time.

18:31 < ShikharJ> If you had to choose two over these four, which ones would you prioritize, if I may ask?

18:33 < zoq> hm, probably Wasserstein GAN and DCGAN, but as I said I don't mind to go with anything else :)

18:35 < ShikharJ> Thanks for sharing your thoughts, I'll plan out my ideas accordingly.

18:36 < zoq> Sounds good, let me know if you have any further questions.

21:56 < jon_> hi

22:02 < zoq> jon_: Hello there!

22:02 < jon_> Is this for mlpack development discussion

22:04 < zoq> yes, here, github, mailinglist works

22:05 < jon_> is there a TSNE implementation in mlpack?

22:06 < zoq> unfortunately no

22:06 < jon_> I don't see one and am wondering if contributing one would be worthwhile

22:09 < zoq> I like the idea, I guess LargeVis might also be interesting.

22:11 < zoq> Also mlpack has an SpillTree implementation, which could be used.

22:33 < travis-ci> mlpack/mlpack#3800 (master - 119c410 : Marcus Edel): The build has errored.

22:33 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/bd3e7af612fc...119c410c56a1

22:33 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/335833928

22:51 < rcurtin> jon_: zoq: I'd love to include T-SNE; there's even a dual-tree algorithm about it

22:52 < rcurtin> http://www.jmlr.org/papers/volume15/vandermaaten14a/vandermaaten14a.pdf

22:52 < rcurtin> it would be possible (although it would require a little thinking) to implement dual-tree t-SNE (and single-tree) inside of mlpack's framework, and I'd be happy to help with it

22:57 < zoq> Definitely a great way to dive into the codebase especially if someone likes to explore the tree code.