verne.freenode.net changed the topic of #mlpack to: http://www.mlpack.org/ -- We don't respond instantly... but we will respond. Give it a few minutes. Or hours. -- Channel logs: http://www.mlpack.org/irc/
govg has joined #mlpack
travis-ci has joined #mlpack
< travis-ci>
ShikharJ/mlpack#132 (AtrousConv - acf54f4 : aarushgupta): The build has errored.
< ShikharJ>
rcurtin: The jenkins build for Layer Normalization PR is seemingly stuck, when can we expect the queue to clear up?
Atharva has joined #mlpack
< zoq>
let's stop/restart the monthly matrix build
Atharva has quit [Quit: Connection closed for inactivity]
< rcurtin>
sure, let me kill it
< rcurtin>
ah, actually, looks like you already did it :)
ImQ009 has joined #mlpack
govg has quit [Ping timeout: 246 seconds]
vivekp has quit [Ping timeout: 240 seconds]
vivekp has joined #mlpack
manish7294 has joined #mlpack
< manish7294>
rcurtin: I gave another shot to both the single KNN pass and multi pass KNN (BoostMetric Implementation) constraints generation implementation. And here's quick insight: Data points: 5000, Single Pass: K=3 -> 272.695secs, K=40->299.084secs. Now for Multi Pass: k=3->7.2961secs, k=40->Like forever (Already been more than 3 hours and still running wit
< manish7294>
h high CPU usage, Probably have to kill the process)
< rcurtin>
manish7294: I haven't had a chance to look into your response yet but it sounds like the MATLAB implementation is horrifyingly slow
< rcurtin>
not sure what is up with the multipass algorithm you are using but I virtually guarantee a good implementation will be way faster than singlepass KNN with k = N - 1
< rcurtin>
the dimensionality of the data you are using will make a difference, so if you are working in 10k dimensions, brute force will typically be faster than trees
< rcurtin>
but I don't think that is the regime you are working in
< manish7294>
rcurtin: On other hand if you look it like this - for eg: let's we have 3 classes with 20,2,10 instances then in multipass we will be calculating KNN for k = 30, 22,12 whereas we can reduce the single pass knn to just run for k =30
< manish7294>
let the least number of instance of a class is 'X' then we just need to run KNN for N - X + K which is similar to atleast one of the pass of multi KNN
manish7294 has left #mlpack []
manish7294 has joined #mlpack
< rcurtin>
manish7294: sorry, I think there is a misunderstanding. if you run multipass KNN (if that's what we should call it), each run you set k = k, not k = n - 1
< rcurtin>
where 'k' is however many impostors or same-labeled neighbors you are looking for
< rcurtin>
also, keep in mind, the interesting regime for BoostMetric will not be 100-point datasets... it will be 100k or 1M or 10M point datasets
< rcurtin>
basically, we should be focusing on seeing how much we can make it scale on a single system with mlpack
< manish7294>
rcurtin: Ahhh! my bad. Tommorow, I will try implementing your idea. Hopefully we will get better results. Thanks!
< rcurtin>
sure, let me know if there is anything else I can clarify :)