def myGridSearch(data,NTreeBounds,NFeaturesBounds):
best_acc = -float('inf')
bestrandomforest = None
class bestValues(object):
t = float('nan')
f = float('nan')
for t in range(NTreeBounds[0],NTreeBounds[1]+NTreeBounds[2],NTreeBounds[2]):
for f in range(NFeaturesBounds[0],NFeaturesBounds[1]+NFeaturesBounds[2],NFeaturesBounds[2]):
randomforest = RandomForest()
randomforest.setNumTrees(int(t))
randomforest.setNumFeatures(int(f))
evaluation = Evaluation(data)
output = output = util.get_buffer_for_predictions()[0]
attRange = Range() # no additional attributes output
outputDistribution = Boolean(False) # we don't want distribution
random = Random(1)
numFolds = min(10,data.numInstances())
evaluation.crossValidateModel(randomforest,data,numFolds,random,[output, attRange, outputDistribution])
acc = evaluation.pctCorrect()
if (acc>best_acc):
bestrandomforest = randomforest
best_acc = acc
bestValues.t = t
bestValues.f = f
print "Best accuracy:", best_acc
print "Best values: NTreeBounds = ", bestValues.t, ", NFeaturesBounds = ", bestValues.f
print "-----------------------------------------"
return bestrandomforest, bestValues.t, bestValues.f, best_acc
def RandomForest_ParamFinder(data):
# possible set for Number of trees
NTreeBounds = [1,20,1]
# possible set for number of features
NFeaturesBounds = [0,20,1]
if (data.numInstances()>10): # grid search does 10-fold cross validation; hence number of samples must be more than 10
gridsearch = GridSearch()
acctag = gridsearch.getEvaluation()
acctag = SelectedTag('ACC',acctag.getTags())
gridsearch.setEvaluation(acctag)
allfilters = AllFilters()
gridsearch.setFilter(allfilters)
gridsearch.setGridIsExtendable(Boolean(True))
randomforest = RandomForest()
gridsearch.setClassifier(randomforest)
gridsearch.setXProperty(String('classifier.numTrees'))
gridsearch.setYProperty(String('classifier.numFeatures'))
gridsearch.setXExpression(String('I'))
gridsearch.setYExpression(String('I'))
gridsearch.setXMin(NTreeBounds[0])
gridsearch.setXMax(NTreeBounds[1])
gridsearch.setXStep(NTreeBounds[2])
gridsearch.setYMin(NFeaturesBounds[0])
gridsearch.setYMax(NFeaturesBounds[1])
gridsearch.setYStep(NFeaturesBounds[2])
gridsearch.setYBase(10)
print "searching for random-forest NumTrees = [", NTreeBounds[0], ",", NTreeBounds[1], "], NumFeatures = [ ", NFeaturesBounds[0], ",", NFeaturesBounds[1], "] ...."
gridsearch.buildClassifier(data)
bestValues = gridsearch.getValues()
# ----------------------- Evaluation
bestrandomforest = RandomForest()
bestrandomforest.setNumTrees(int(bestValues.x))
bestrandomforest.setNumFeatures(int(bestValues.y))
evaluation = Evaluation(data)
output = output = util.get_buffer_for_predictions()[0]
attRange = Range() # no additional attributes output
outputDistribution = Boolean(False) # we don't want distribution
random = Random(1)
numFolds = min(10,data.numInstances())
evaluation.crossValidateModel(bestrandomforest,data,numFolds,random,[output, attRange, outputDistribution])
acc = evaluation.pctCorrect()
print "best accuracy: ", acc
print "best random-forest classifier with NumTrees=",bestValues.x , ", NumFeatures = ", bestValues.y
OptRndFrst = bestrandomforest
OptRndFrstp1 = bestValues.x
OptRndFrstp2 = bestValues.y
OptRndFrstAcc = acc
else:
OptRndFrst, OptRndFrstp1, OptRndFrstp2, OptRndFrstAcc = myGridSearch(data,NTreeBounds,NFeaturesBounds)
Description = 'Random-Forest classifier: OptNumTrees = ' + str(OptRndFrstp1) + \
', OptNumFeatures = ' + str(OptRndFrstp2) + ', OptAcc = ' + str(OptRndFrstAcc)
print "-----------------------------------------"
return OptRndFrst, OptRndFrstp1, OptRndFrstp2, OptRndFrstAcc, Description
def random_forest(trainData,testData,params,exparams):
numTrees = int(float(params[0]))
numFeatures = int(float(params[1]))
randomforest = RandomForest()
randomforest.setNumTrees(numTrees)
randomforest.setNumFeatures(numFeatures)
randomforest.buildClassifier(trainData) # only a trained classifier can be evaluated
# evaluate it on the training
evaluation = Evaluation(trainData)
(trainOutput, trainBuffer) = util.get_buffer_for_predictions(trainData)
attRange = Range() # no additional attributes output
outputDistribution = Boolean(False) # we don't want distribution
evaluation.evaluateModel(randomforest, trainData, [trainOutput, attRange, outputDistribution])
print "--> Evaluation:\n"
print evaluation.toSummaryString()
trainSummary = makeTrainEvalSummary(evaluation)
# evaluate it on testing
evaluation = Evaluation(testData)
(testOutput, testBuffer) = util.get_buffer_for_predictions(testData)
attRange = Range() # no additional attributes output
outputDistribution = Boolean(False) # we don't want distribution
evaluation.evaluateModel(randomforest, testData, [testOutput, attRange, outputDistribution])
return trainBuffer, testBuffer, trainSummary