Hi ! Thanks a lot for sending the draft! Great initiative. Looks nice, and this is definitely something useful. There are many details that we would need to work on in order to add this notebook to the repo, but that is not too important for now. I suggest you polish it a bit and send a PR (not urgent!)
As you have showed us some nice work regarding using Shogun's Python API and understanding the principle of approximate kernel feature embeddings, I would suggest that you move your attention to sending in some project related c++ code now. We haven't seen a lot of code from you yet, and just to repeat the advice we give all students: if you want to do a c++ projects, you should send a PR with some nontrivial c++ code. In your case, think any of * clean up the existing RFF code. Examples looking at the RandomFourierGaussPreproc.cpp: -remove SG_MALLOC, refactor using SGVector/SGMatrix -improve the error messages -merge cur_kernelwidth and kernelwidth, and same for the other vars -remove "dim_input_space" which shouldnt be needed -remove test_rfinited, which is really bad design -dont define PI locally, use the one in C++ -use our gaussian rng rather than doing the gauss transform from uniform numbers (lines 246ff) -performance improve: don't apply the embedding vector wise when receiving a matrix. Mat-Mat products are faster! You see there is lots to do :) Furthermore, you can show us some code design skills: * merge the two redundant implementations of RFF in transformer and features so that they share code * add a spectral density for another kernel, i.e. random features for something else than a Gaussian kernel. This is mostly design work: how to extend what we have to other kernels? * Draft the Nystrom based embedding or FastFood. * Implement a C++ version of the incomplete Cholesky factorization. Am So., 5. Apr. 2020 um 21:56 Uhr schrieb sai_ng via shogun-list < shogun-list@shogun-toolbox.org>: > Hey there, > I've a proposal, I made an explicit feature mapping notebook as part of my > GSOC proposal. I really believe that it would be beneficial if it's under > our ```docs/ipython-notebook/preprocessors``` so I'm sending the gist > before I make a PR on it. There are still some features which aren't > working in all honesty, but nothing that can't be fixed :D !. Looking > forward to your feedback. > > https://gist.github.com/NanuSai/cec79b5de5ebc3a90b7889843a8bfaa5 >
