2010/1/24 Ted Dunning <ted.dunn...@gmail.com>: > The other take-away was that threading looks plausible for SGD, but full on > map-reduce except on somewhat randomized shards of features probably isn't > that useful. Even shard may not be very useful since different mappers (or > reducers) may just mostly redo the same work.
Yes, I had the same overall feeling about the mapreducability of the training. Good threading/multicore support sounds more like a good mid-term objective for online learners in Mahout. MapReduce is still very interesting for feature extraction for large training / testing datasets (hashing text is still CPU intensive and completely parallelizable). > On Sun, Jan 24, 2010 at 5:37 AM, Olivier Grisel > <olivier.gri...@ensta.org>wrote: > >> The main takeway point is that averaging for linear models is possible >> but not as interesting as horizontal feature sharding that >> experimentally works for both linear and non linear models. >> >> The second takeway point is that vowpal wabbit looks more and more >> unbeatable :) > > Unbeatable perhaps. But that also makes it pretty important to have in > Mahout. Yes sure. Speaking of which I did some more work to wrap the online logistic regression model into a multi label document classifier: http://github.com/ogrisel/mahout/commits/MAHOUT-228 A simple unittest with a toy dataset of around 20 sentences categorized in 0 to 3 categories confirms the models converges towards good F1 measure. I will now work on larger dataset in the example package based on the wikipedia extractor. -- Olivier http://twitter.com/ogrisel - http://code.oliviergrisel.name
