speed up t-SNE by using BLAS.syrk() for Q=YY^T and calculating only the upper triangle of the symmetric matrices (MNIST test times go from 33-36 secs to 23-26 secs on my laptop)
replace FactCheck with the built-in @testset (requires Julia 0.5, BaseTestNext used for backward compatibility with 0.4)
cleanup examples code
edit README
I can split the method improvements and the examples/documentation changes into separate PRs, if you would like to review it separately.
Coverage increased (+0.07%) to 99.296% when pulling da2e8e186d2dcebb39b6792643bbf2aaa17039d6 on alyst:enh_perf into d71ef231db06c9b38f38ac412706fde2ec9c0ddf on lejon:master.
Coverage increased (+0.07%) to 99.296% when pulling bd9611ba96376513e40b9e46ab62677339d7aee9 on alyst:enh_perf into d71ef231db06c9b38f38ac412706fde2ec9c0ddf on lejon:master.
Coverage increased (+0.07%) to 99.296% when pulling bd9611ba96376513e40b9e46ab62677339d7aee9 on alyst:enh_perf into d71ef231db06c9b38f38ac412706fde2ec9c0ddf on lejon:master.
Coverage increased (+0.07%) to 99.296% when pulling bd9611ba96376513e40b9e46ab62677339d7aee9 on alyst:enh_perf into d71ef231db06c9b38f38ac412706fde2ec9c0ddf on lejon:master.
It's me again with some changes:
BLAS.syrk()
forQ=YY^T
and calculating only the upper triangle of the symmetric matrices (MNIST test times go from 33-36 secs to 23-26 secs on my laptop)FactCheck
with the built-in@testset
(requires Julia 0.5,BaseTestNext
used for backward compatibility with 0.4)I can split the method improvements and the examples/documentation changes into separate PRs, if you would like to review it separately.