def testSingle(exampleArr,outcomeArr,theta,muArr,sigmaArr):
"""
tests that classification of training array using hypothesis(theta)
produces the same outcomes as the outcome array
"""
# don't you have to normalise this 1st?
outcomeCalcRaw = np.dot(exampleArr,theta)
outcomeCalc = logReg.sigmoid(outcomeCalcRaw)
outcomeCalc[outcomeCalc>0.5] = 1
outcomeCalc[outcomeCalc<=0.5] = 0
numTrainEx = float(exampleArr.shape[0])
correct = 100*(outcomeCalc==outcomeArr).sum()/numTrainEx
return correct
def testAll(dataType):
"""
tests one vs all classification over all categories
"""
# load all theta values
readName = dataType+'Theta.csv'
(numCat,numFeat,dataType,catNames,theta,mu,sigma) = loadResults(readName,loadForClassify=True)
n = numFeat
print theta.shape,mu.shape
for i in range(numCat):
# load labelled digit data
saveName = dataType+catNames[i]+'Data.csv'
digitData = np.genfromtxt(saveName, delimiter=',')
m = digitData.shape[0]
# get X, y from data
X = digitData[:,:(n)]
y = digitData[:,(n)]
# add Xo = 1 into X
Xcalc = np.zeros((m,n+1))
Xcalc[:,1:] = X.copy()
Xcalc[:,0] = 1
# normalise X
Xcalc = (Xcalc - mu[:,i]) / sigma[:,i]
# generate hypothesis for each training example
yCalcArr = np.zeros((m,numCat))
for j in range(numCat):
yCalcRaw = np.dot(Xcalc,theta[:,j])
yCalcArr[:,j] = logReg.sigmoid(yCalcRaw)
yCalc = np.zeros(m)
for j in range(m):
bestEst = np.where(yCalcArr[j,:]==max(yCalcArr[j,:]))[0][0]
if bestEst == i:
yCalc[j] = 1
else:
yCalc[j] = 0
# compare calc Y to empirical Y
correct = 100*(yCalc==y).sum()/float(m)
print 'category',catNames[i]
print 'hypothesis generated by logistic regression produces ',correct,' percent \ncorrect classifications of training data'
def classifyImage(imgArr,imgType,parameterName):
"""
does one vs all classification only
currently only works for digits
will have to add suits too
consider adding this info to parameter files so that this can be automated more simply
"""
imageSpecs = ImageSpecs(imgType)
#numCol = imageSpecs.numCol
#numRow = imageSpecs.numRow
# load hypothesis function (theta values)
path = os.path.normpath('learningData/'+parameterName)
(numCat,numFeat,dataType,catNames,theta,mu,sigma) = loadResults(path,loadForClassify=True)
# get X (input data for classfication) from image data
#X = buildTrainingExampleArray(imgArr,dataType,numCol,numRow,isRGB=False)
X = buildTrainingExampleArray(imgArr,imageSpecs,isRGB=False)
# add Xo = 1 into X
Xcalc = np.zeros(numFeat)
Xcalc[1:] = X.copy()
Xcalc[0] = 1
yCalcArr = np.zeros(numCat)
# generate hypothesis for each digit
for i in range(numCat):
# normalise X
XcalcCat = (Xcalc - mu[:,i]) / sigma[:,i]
# generate hypothesis
yCalcRaw = np.dot(XcalcCat,theta[:,i])
yCalcArr[i] = logReg.sigmoid(yCalcRaw)
bestEst = np.where(yCalcArr==yCalcArr.max())[0][0]
catEst = catNames[bestEst]
#print yCalcArr
#print 'best estimate index',bestEst,'classifies digit as',catEst
return catEst