wrapped all the randomizers in training.clj

src/clojure_encog/examples.clj

@@ -20,10 +20,10 @@
       trainer   ((make-trainer :resilient-prop) network dataset)]
      ;;make use of the boolean parameter
       (if-not keep-trying? 
-              (train trainer 0.01 300 #_(RequiredImprovementStrategy. 5)) ;;train the network once
+              (train trainer 0.01 300 [] #_[(RequiredImprovementStrategy. 5)]) ;;train the network once
       (loop [t false counter 0 _ nil] 
       (if t (println "Nailed it after" (str counter) "times!")
-      (recur  (train trainer 0.01 300 #_(RequiredImprovementStrategy. 5))  ;;train the network until it succeeds
+      (recur  (train trainer 0.01 300 [] #_[(RequiredImprovementStrategy. 5)])  ;;train the network until it succeeds
               (inc counter) (. network reset)))) )     
 (do (println "\nNeural Network Results:")
     (doseq [pair dataset] 

src/clojure_encog/training.clj

@@ -18,7 +18,11 @@
        (org.encog.ml.svm.training SVMTrain)
        (org.encog.ml.svm SVM)
        (org.encog.ml MLRegression)
+       (org.encog.util.simple EncogUtility)
        (org.encog.neural.networks BasicNetwork)
+       (org.encog.mathutil.randomize Randomizer BasicRandomizer ConsistentRandomizer 
+                                     ConstRandomizer FanInRandomizer Distort GaussianRandomizer
+                                     NguyenWidrowRandomizer RangeRandomizer)
        (org.encog.util.arrayutil NormalizeArray TemporalWindowArray)))
 
 
@@ -43,6 +47,26 @@
 :else (throw (IllegalArgumentException. "Unsupported data model!"))
 ))
 
+(defn make-randomizer 
+"Constructs a Randomizer object. "
+[type]
+(condp = type
+    :basic      (BasicRandomizer.) ;random number generator with a random(current time) seed
+    :range      (fn [min-val max-val] (RangeRandomizer. min-val max-val)) ;range randomizer
+    :consistent (fn [min-rand max-rand] (ConsistentRandomizer. min-rand max-rand)) ;consistent range randomizer
+    :constant   (fn [constant] (ConstRandomizer. constant))
+    :distort    (fn [factor]   (Distort. factor))
+    :fan-in     (fn [boundary sqr-root?] (FanInRandomizer. (unchecked-negate boundary) boundary (if (nil? sqr-root?) false sqr-root?)))
+    :gaussian   (fn [mean st-deviation] (GaussianRandomizer. mean st-deviation))
+    :nguyen-widrow  (NguyenWidrowRandomizer.)   
+))
+
+(defn randomize
+"Performs the actual randomization. Expects a randomizer object and some data. Options for data include:
+ MLMethod -- double -- double[] -- double[][] -- Matrix " 
+[^Randomizer randomizer data] 
+(. randomizer randomize data))
+
 (defmacro judge 
 "Consumer convenience for implementing the CalculateScore interface which is needed for genetic and simulated annealing training."
 [minimize? & body]
@@ -51,7 +75,13 @@
   (shouldMinimize [] ~minimize?)))
 
 (defn make-trainer
-"Constructs a training-method object."
+"Constructs a training-method object given a method. Options inlude:
+-------------------------------------------------------------
+ :simple     :back-prop    :quick-prop      :manhattan         
+ :genetic    :svm          :nelder-mead     :annealing     
+ :scaled-conjugent         :resilient-prop  :pnn                
+ ------------------------------------------------------------- 
+ Returns a MLTrain object."
 [method]
 (condp = method
        :simple     (fn [net tr-set learn-rate] (TrainAdaline.  net tr-set (if (nil? learn-rate) 2.0 learn-rate)))
@@ -85,18 +115,28 @@
 
 
 (defn train 
-"Does the actual training. This is a potentially lengthy and costly process so most type hints have been provided. Returns true or false depending on whether the error target was met within the iteration limit."
-[^MLTrain method ^Double error-tolerance ^Integer limit & strategies] ;;eg: (new RequiredImprovementStrategy 5)
+"Does the actual training. This is a potentially lengthy and costly process so most type hints have been provided. Returns true or false depending on whether the error target was met within the iteration limit. This is an overloaded fucntion. It is up to you whether you want to provide limits for error-tolerance, iteration-number or both."
+([^MLTrain method ^Double error-tolerance ^Integer limit strategies] ;;eg: (new RequiredImprovementStrategy 5)
 (when (seq strategies) (dotimes [i (count strategies)] 
                        (.addStrategy method (nth strategies i))))
-(loop [epoch (int 1)]
-(if (< limit epoch) false ;;failed to converge
-(do (. method iteration)
-    (println "Epoch #" epoch " Error:" (. method getError))    
-(if-not (> (. method getError) 
-           error-tolerance) true ;;succeeded to converge 
-(recur (inc epoch)))))))
+     (loop [epoch (int 1)]
+       (if (< limit epoch) false ;;failed to converge
+       (do (. method iteration)
+           (println "Epoch #" epoch " Error:" (. method getError))    
+       (if-not (> (. method getError) error-tolerance) true ;;succeeded to converge 
+       (recur (inc epoch)))))))
+
+([^MLTrain method ^Double error-tolerance strategies] 
+(when (seq strategies) (dotimes [i (count strategies)] 
+                       (.addStrategy method (nth strategies i))))
+     (EncogUtility/trainToError method error-tolerance))
 
+([^MLTrain method strategies] ;;requires only one iteration - SVMs or Nelder-Mead for example
+ (when (seq strategies) (dotimes [i (count strategies)] 
+                        (.addStrategy method (nth strategies i))))
+     (do (. method iteration)
+         (println "Error:" (. method getError)))))
+
 
 (defn normalize 
 "Normalizes a seq (vector/list) of doubles (no nests) between high-end low-end and returns the normalized double array. Call seq on the result to convert it back to a clojure seq so it reads nicely on the repl." 
@@ -105,7 +145,8 @@
 (do (. norm setNormalizedHigh high-end)
     (. norm setNormalizedLow  low-end)  
     (. norm process (double-array data)))))
-
+
+