Laurens van der Maaten

GPUs tend to be memory-bound rather than compute-bound, in particular, for small models that require additional memory transfers such as ShuffleNets and CondenseNets. On mobile devices, embedded systems, etc. the...

There are basically two important loops that should be straightforward to parallelize: https://github.com/lvdmaaten/bhtsne/blob/master/sptree.cpp#L385 https://github.com/lvdmaaten/bhtsne/blob/master/tsne.cpp#L215 The first loop is actually embarrassingly parallel, so it should be completely trivial. The second loop...

I'm no OpenMP expert, but it looks good to me. What kind of speed-ups are you seeing compared to the non-OpenMP code?

Nice! In an earlier version of the code, I had hardcoded the output dimension. Indeed, that was a lot faster than the version that is currently in the repo.

Nice!

By definition, the loss of any multidimensional scaling method (that is, a loss that measures to what extent pairwise distances are preserved) is invariant under rotations and inflections of the...

There is some discussion on that here: https://github.com/lvdmaaten/bhtsne/issues/7 It should be straightforward to add additional distance measures, as long as these are metric: that is, as long as they satisfy...

This is not implemented right now. Both Matlab and Numpy support compressed sparse row matrices, so it would be possible to add this. For now, a potential way to circumvent...

Is the t-SNE algorithm itself actually being run? Like do you see a loss being printed every 50 iterations or something like that? The result file should actually be really...

Can you please copy-paste the full output? Does the `data.dat` file get written by the Python wrapper?