@@ -35,7 +35,7 @@ (defn from-reader "Returns a lazy sequence of examples parsed from the given Reader" [reader] - (map parse (-> reader BufferedReader. line-seq) (iterate inc 1))) + (pmap parse (-> reader BufferedReader. line-seq) (iterate inc 1))) ;; --- Small RCV Data Set - 2000 examples ---- (def rcv-small-file "../data/train2000.dat.gz") online/data.clj @@ -6,7 +6,7 @@ (ns loss (:use vec (clojure.contrib profile))) -(def *max-deriv* 100000) +(def *max-deriv* 10) (defn deriv "Returns the (clamped) derivative of the loss l evaluated at y v" @@ -26,7 +26,7 @@ (defn exp ([y v] (prof :exp (Math/pow Math/E (- (* y v))))) - ([y v _] (prof :dexp (- (* y (exp y v)))))) + ([y v _] (prof :dexp (* (- y) (exp y v))))) (defn logistic ([y v] (prof :logistic (Math/log (+ 1 (exp y v))))) online/loss.clj @@ -1,3 +1,14 @@ +Using Different Losses +====================== + +It seems hinge loss outperforms logistic as a surrogate for 0-1 loss. +Error rates for full training set and initial 1000 as test set: + +* Logistic: 6.1% +* Hinge: 3.9% + +(Notes: Projection frequency = 1, Max. derivative = 10) + Profiling Notes =============== online/notes.md @@ -22,25 +22,23 @@ ; Notes ; ----- -; * Parsing all of the examples in the full 781,000+ data set takes 2:40 on my -; 2.66GHz 2Gb RAM iMac ; ; * Memory uses rises quickly to around 160Mb of real memory then stablises. ; ; To Do ; ----- -; * Optimise gradient calculations in SGD via rearranging calls to add & scale. +; [_] Improve the parsing speed so that lines are parsed char-by-char into maps. ; -; * Improve the parsing speed so that lines are parsed char-by-char into maps. +; [X] Optimise gradient calculations in SGD via rearranging calls to add & scale. ; -; * Make the model vector in SGD dense and update vector methods to handle +; [X] Make the model vector in SGD dense and update vector methods to handle ; sparse/dense updates. (ns run (:use (clojure.contrib profile))) (ns clojure.contrib.profile) -(def *enable-profiling* true) +(def *enable-profiling* false) (ns run (:require data learner sgd loss)) @@ -51,7 +49,7 @@ (def *loss* loss/hinge) (def *lambda* 0.0001) -(def *projection-freq* 10) +(def *projection-freq* 1) (time ; (profile @@ -59,4 +57,4 @@ (sgd/make-learner *loss* *lambda* *projection-freq*) (take (+ *num-train* *num-test*) (data/stdin)) *num-test* - *report-freq*)) ;) + *report-freq*));) online/run.clj fa89732a9805bbbae2ae05a5d7dc16a02f7e29fa mreid mark@reid.name http://github.com/mreid/injuce/commit/60a5a78696b195868e2712bc98058b915ddf6017 60a5a78696b195868e2712bc98058b915ddf6017 2009-10-09T03:00:14-07:00 2009-10-09T03:00:14-07:00 Parsing now uses pmap for a 33% speed boost on my dual core machine. 00ca87a071636814e6b909ab7d48838ae02608de mreid mark@reid.name