def test(self, categories):
for i in range(Constants.NUM_SUBJECTS):
trainSubjects = [1, 2, 3, 4]
testSubjects = [i + 1]
trainSubjects.remove(i + 1)
trainVoxelArrayMap = util.getVoxelArray(subjectNumbers = trainSubjects)
testVoxelArrayMap = util.getVoxelArray(subjectNumbers = testSubjects)
util.normalize(trainVoxelArrayMap)
util.normalize(testVoxelArrayMap)
util.filterData(trainVoxelArrayMap, categories=categories)
util.filterData(testVoxelArrayMap, categories=categories)
Xtrain = numpy.array([trainVoxelArrayMap[key] for key in trainVoxelArrayMap])
Ytrain = numpy.array([key[1] for key in trainVoxelArrayMap])
Xtest = numpy.array([testVoxelArrayMap[key] for key in testVoxelArrayMap])
Yanswer = numpy.array([key[1] for key in testVoxelArrayMap])
Yprediction = OneVsRestClassifier(LinearSVC()).fit(Xtrain, Ytrain).predict(Xtest)
# Yprediction = OneVsOneClassifier(LinearSVC()).fit(Xtrain, Ytrain).predict(Xtest)
correct = 0
for index in range(len(Yanswer)):
if Yanswer[index] == Yprediction[index]:
correct += 1
# correct = [1 if Yanswer[index] == Yprediction[index] else 0 for index in range(len(Yanswer))]
print categories, "Correct Predictions: ", correct, "/", len(Yanswer)
return float(correct) * 100 / len(Yanswer)
def correlationNearestNeighbor():
voxelArrayMap = util.getVoxelArray(False, False,True,False, [4])
util.normalize(voxelArrayMap)
correct = [{ }]*3
totalCountCorrect = [0] *3
totalCountInCorrect = [0] *3
incorrect = [{ }]*3
voxelCopy = voxelArrayMap.keys()
totalCorrect =0
totalIncorrect =0
count = 0
for key in voxelCopy:
count +=1
testExample = voxelArrayMap[key]
voxelArrayMap.pop(key, None)
averageCategoryCorrelations = matrixify(calculateAverageCorrelations(voxelArrayMap))
exampleCorrelations = calculateSingleExampleCorrelations(testExample, voxelArrayMap)
classifiedCategory = classifyByClosestCorrelation(exampleCorrelations,averageCategoryCorrelations)
if classifiedCategory[0] == key[1]:
totalCorrect +=1
else:
totalIncorrect +=1
voxelArrayMap[key] = testExample
print "Correct", totalCorrect, "Incorrect", totalIncorrect, "Percentage Correct ", totalCorrect/float((totalCorrect +totalIncorrect))
def __init__(self, data=None, normalize=True, numTrainingExamples=24, categories=[0, 1, 2, 3, 4], subjects=[1, 2, 3, 4]):
if data is None:
data = util.getVoxelArray(subjectNumbers=subjects)
if normalize:
util.normalize(data)
if len(categories) < Constants.NUM_CATEGORIES:
util.filterData(data, categories)
self.numCategories = Constants.NUM_CATEGORIES
self.categories = None
if not categories is None:
self.numCategories = len(categories)
self.categories = categories
self.numTrainingExamples = numTrainingExamples * len(subjects) / Constants.NUM_SUBJECTS
self.numTestingExamples = len(subjects) * Constants.NUM_RUNS * Constants.NUM_TASKS - self.numTrainingExamples
self.trainExamples = [[[0] for _ in range(self.numTrainingExamples)] for _ in range(self.numCategories)]
self.testExamples = []
self.averages = [[0 for _ in range(Constants.NUM_VOXELS)] for _ in range(self.numCategories)]
for key in data:
value = data[key]
run, category, task, subject = key
if not self.categories is None:
if category not in self.categories: continue
category = self.getCategory(category)
index = self.index(run, task, subject) * len(subjects) / Constants.NUM_SUBJECTS
if index < self.numTrainingExamples:
self.trainExamples[category][index] = list(value)
else:
self.testExamples.append((key, value))
def run(categories=None, subject=None): data = util.getVoxelArray(subjectNumbers=subject) # data = util.dataSmoothing(data) nb = NaiveBayes(data=data, categories=categories, subjects=subject) nb.train() nb.test() return nb.results()
def correlationNearestNeighbor():
voxelArrayMap = util.getVoxelArray(False, False, True, False, [4])
util.normalize(voxelArrayMap)
correct = [{}] * 3
totalCountCorrect = [0] * 3
totalCountInCorrect = [0] * 3
incorrect = [{}] * 3
voxelCopy = voxelArrayMap.keys()
totalCorrect = 0
totalIncorrect = 0
count = 0
for key in voxelCopy:
count += 1
testExample = voxelArrayMap[key]
voxelArrayMap.pop(key, None)
averageCategoryCorrelations = matrixify(
calculateAverageCorrelations(voxelArrayMap))
exampleCorrelations = calculateSingleExampleCorrelations(
testExample, voxelArrayMap)
classifiedCategory = classifyByClosestCorrelation(
exampleCorrelations, averageCategoryCorrelations)
if classifiedCategory[0] == key[1]:
totalCorrect += 1
else:
totalIncorrect += 1
voxelArrayMap[key] = testExample
print "Correct", totalCorrect, "Incorrect", totalIncorrect, "Percentage Correct ", totalCorrect / float(
(totalCorrect + totalIncorrect))
def __init__(self, data=None, normalize=True, categories=[0, 1, 2, 3, 4], subjects=[1, 2, 3, 4]):
if data is None:
data = util.getVoxelArray(subjectNumbers=subjects)
if normalize:
util.normalize(data)
if len(categories) < Constants.NUM_CATEGORIES:
util.filterData(data, categories)
self.data = data
self.categories = categories
def run():
voxelArrayMap = util.getVoxelArray(False, False, False, True, [2])
util.normalize(voxelArrayMap)
centroids = [0] * 5
keys = list(voxelArrayMap.keys())
selected = []
for index in range(len(centroids)):
while True:
centroidKey = random.choice(keys)
category = centroidKey[1]
if category not in selected:
selected.append(category)
break
print centroidKey
centroids[index] = voxelArrayMap[centroidKey]
numIters = 120
assignmentMap = {}
for iteration in range(numIters):
newCentroids = [ [0] * 1973 for i in range(len(centroids))]
clusterCounts = [0] * len(newCentroids)
for key in voxelArrayMap:
value = voxelArrayMap[key]
distances = [ util.getDistance(centroid, value) for centroid in centroids ]
minDistance = min(distances)
optCentroid = distances.index(minDistance)
assignmentMap[key] = optCentroid
for voxelIndex in range(len(newCentroids[optCentroid])):
newCentroids[optCentroid][voxelIndex] += value[voxelIndex]
clusterCounts[optCentroid] += 1
for index in range(len(newCentroids)):
numPoints = clusterCounts[index]
if numPoints != 0:
centroid = newCentroids[index]
centroid = [ centroid[voxelIndex] / numPoints for voxelIndex in range(len(centroid)) ]
newCentroids[index] = centroid
centroids = newCentroids
# we want ClusterNum -> What type of image
catMap = {}
for key in assignmentMap:
assignment = assignmentMap[key]
# key[1] is the category (0-5 for face, body, etc)
# assignment is the optimal centroid
if assignment in catMap:
arr = catMap[assignment]
arr.append(key[1])
catMap[assignment] = arr
else:
catMap[assignment] = [key[1]]
print catMap
def PCA(normalizedData=None):
if normalizedData is None:
normalizedData = util.getVoxelArray(allTasks=True)
# prepocess and normalize data
util.normalize(normalizedData)
m = len(normalizedData)
print('Calculating the Covariance Matrix')
count = 0
# compute Covariance matrix [1/m Sum (x xT)]
covarianceMatrix = [[0 for _ in range(Constants.NUM_VOXELS)]
for _ in range(Constants.NUM_VOXELS)]
for key in normalizedData:
x = normalizedData[key]
for i in range(Constants.NUM_VOXELS):
for j in range(Constants.NUM_VOXELS):
if count % 50000000 is 0:
print(
int(
float(count) * 100 /
(m * Constants.NUM_VOXELS * Constants.NUM_VOXELS)),
'percent completed')
covarianceMatrix[i][j] += (x[i] * x[j]) / m
count += 1
for i in range(Constants.NUM_VOXELS):
for j in range(Constants.NUM_VOXELS):
if isinstance(covarianceMatrix[i][j], complex):
print('Covariance Matrix index', i, j, 'is complex:',
covarianceMatrix[i][j])
print('Finding the', Constants.K_DIMENSIONS, 'Largest Eigenvalues')
# then compute all eigenvectors and corresponding eigenvalues
eigenvalues, eigenvectors = linalg.eig(covarianceMatrix)
eigens = zip(eigenvalues, eigenvectors)
eigens.sort(key=lambda x: x[0], reverse=True)
eigens = eigens[0:Constants.K_DIMENSIONS]
eigenvalues, eigenvectors = zip(*eigens)
for i in range(Constants.K_DIMENSIONS):
print(i, eigenvalues[i])
# transform data into K dimensions
for key in normalizedData:
x = normalizedData[key]
y = [sum([eigenvector[i] * x[i] for i in range(Constants.NUM_VOXELS)])\
for index, eigenvector in enumerate(eigenvectors)]
normalizedData[key] = y
pylab.show()
def twoDHeatMap(self,
data,
vmin=None,
vmax=None,
title=None,
xLabel=None,
yLabel=None,
filename=None):
if not vmin is None and not vmax is None:
pylab.pcolor(numpy.array(data), vmin=vmin, vmax=vmax)
else:
pylab.pcolor(numpy.array(data))
if not title is None:
pylab.title(title)
if not xLabel is None:
pylab.xlabel(xLabel)
if not yLabel is None:
pylab.ylabel(yLabel)
if not filename is None:
pylab.savefig(filename + '.png')
pylab.show()
if __name__ == '__main__':
data = util.getVoxelArray()
data = util.dataSmoothing(data)
util.normalize(data)
g = Grapher()
g.twoDGraph(data)
def test(firstCategory, secondCategory):
def sigmoid(x, deriv=False):
return x * (1 - x) if (deriv == True) else 1 / (1 + numpy.exp(-x))
trainVoxelMap = util.getVoxelArray(allTasks=True,
includeOddBall=False,
includeWorkingMemory=False,
includeSelectiveAttention=True,
subjectNumbers=trainSubjects)
testVoxelMap = util.getVoxelArray(allTasks=True,
includeOddBall=False,
includeWorkingMemory=False,
includeSelectiveAttention=True,
subjectNumbers=testSubjects)
util.normalize(trainVoxelMap)
util.normalize(testVoxelMap)
trainX = []
trainY = []
testX = []
testY = []
category1 = firstCategory
category2 = secondCategory
def testWithinSubjects(): # separate by runs
for key in trainVoxelMap:
if key[1] == category1 or key[1] == category2:
classification = 1 if key[1] == category2 else 0
if key[0] == 2:
testX.append(trainVoxelMap[key])
testY.append(classification)
else:
trainX.append(trainVoxelMap[key])
trainY.append(classification)
def testAcrossSubjects(): # include all data for a whole subject
for key in trainVoxelMap:
if key[1] == category1 or key[1] == category2:
trainX.append(trainVoxelMap[key])
classification = 1 if key[1] == category2 else 0
trainY.append(classification)
for key in testVoxelMap:
if key[1] == category1 or key[1] == category2:
testX.append(testVoxelMap[key])
classification = 1 if key[1] == category2 else 0
testY.append(classification)
testWithinSubjects()
trainX = numpy.array(trainX)
trainY = numpy.array(trainY).T
# back propogation
numpy.random.seed(1)
synapseZero = 2 * numpy.random.random((1973, 1)) - 1
for iter in xrange(10000):
l0 = trainX
l1 = sigmoid(numpy.dot(l0, synapseZero))
l1_error = [trainY[i] - l1[i] for i in range(len(trainY))]
l1_delta = l1_error * sigmoid(l1, True)
synapseZero += numpy.dot(l0.T, l1_delta)
#print "Correct Classifications: ", testY
a = sigmoid(numpy.dot(testX, synapseZero))
predictions = [
0 if abs(a[i]) < abs(a[i] - 1) else 1 for i in range(len(testX))
]
#print "Output After Training: ", predictions
print "Categories: ", category1, "/", category2, " Error: ", sum([
1 if predictions[i] != testY[i] else 0 for i in range(len(testY))
]), " out of ", len(testY)
def test(firstCategory, secondCategory): def sigmoid(x,deriv=False): return x*(1-x) if(deriv==True) else 1/(1+numpy.exp(-x)) trainVoxelMap = util.getVoxelArray(allTasks=True, includeOddBall=False, includeWorkingMemory=False, includeSelectiveAttention=True, subjectNumbers = trainSubjects) testVoxelMap = util.getVoxelArray(allTasks=True, includeOddBall=False, includeWorkingMemory=False, includeSelectiveAttention=True, subjectNumbers = testSubjects) util.normalize(trainVoxelMap) util.normalize(testVoxelMap) trainX = [] trainY = [] testX = [] testY = [] category1 = firstCategory category2 = secondCategory def testWithinSubjects(): # separate by runs for key in trainVoxelMap: if key[1] == category1 or key[1] == category2: classification = 1 if key[1] == category2 else 0 if key[0] == 2: testX.append(trainVoxelMap[key]) testY.append(classification) else: trainX.append(trainVoxelMap[key]) trainY.append(classification) def testAcrossSubjects(): # include all data for a whole subject for key in trainVoxelMap: if key[1] == category1 or key[1] == category2: trainX.append(trainVoxelMap[key]) classification = 1 if key[1] == category2 else 0 trainY.append(classification) for key in testVoxelMap: if key[1] == category1 or key[1] == category2: testX.append(testVoxelMap[key]) classification = 1 if key[1] == category2 else 0 testY.append(classification) testWithinSubjects() trainX = numpy.array(trainX) trainY = numpy.array(trainY).T # back propogation numpy.random.seed(1) synapseZero = 2*numpy.random.random((1973, 1)) - 1 for iter in xrange(10000): l0 = trainX l1 = sigmoid(numpy.dot(l0,synapseZero)) l1_error = [ trainY[i]-l1[i] for i in range(len(trainY)) ] l1_delta = l1_error * sigmoid(l1, True) synapseZero += numpy.dot(l0.T, l1_delta) #print "Correct Classifications: ", testY a = sigmoid(numpy.dot(testX,synapseZero)) predictions = [0 if abs(a[i]) < abs(a[i] - 1) else 1 for i in range(len(testX))] #print "Output After Training: ", predictions print "Categories: ", category1, "/", category2, " Error: ", sum([1 if predictions[i] != testY[i] else 0 for i in range(len(testY))]), " out of ", len(testY)
def run(categories, subjects): data = util.getVoxelArray(subjectNumbers=subjects) nn = NearestNeighbor(data=data, categories=categories) return nn.kNearestNeighbor(5)
def test(firstCategory, secondCategory):
def sigmoid(x, deriv=False):
return x * (1 - x) if (deriv == True) else 1 / (1 + numpy.exp(-x))
# initialization
trainVoxelMap = util.getVoxelArray(allTasks=True,
includeOddBall=False,
includeWorkingMemory=False,
includeSelectiveAttention=True,
subjectNumbers=trainSubjects)
testVoxelMap = util.getVoxelArray(allTasks=True,
includeOddBall=False,
includeWorkingMemory=False,
includeSelectiveAttention=True,
subjectNumbers=testSubjects)
util.normalize(trainVoxelMap)
util.normalize(testVoxelMap)
trainX = []
trainY = []
testX = []
testY = []
category1 = firstCategory
category2 = secondCategory
def testWithinSubjects():
for key in trainVoxelMap:
if key[1] == category1 or key[1] == category2:
classification = 1 if key[1] == category2 else 0
if key[0] == 2:
testX.append(trainVoxelMap[key])
testY.append(classification)
else:
trainX.append(trainVoxelMap[key])
trainY.append(classification)
def testAcrossSubjects():
for key in trainVoxelMap:
if key[1] == category1 or key[1] == category2:
trainX.append(trainVoxelMap[key])
classification = 1 if key[1] == category2 else 0
trainY.append(classification)
for key in testVoxelMap:
if key[1] == category1 or key[1] == category2:
testX.append(testVoxelMap[key])
classification = 1 if key[1] == category2 else 0
testY.append(classification)
testWithinSubjects()
trainX = numpy.array(trainX)
trainY = numpy.array(trainY).T
numpy.random.seed(1)
synapseZero = 2 * numpy.random.random((1973, 45)) - 1
synapseOne = 2 * numpy.random.random((45, 1)) - 1
for iter in range(5000):
l0 = trainX
l1 = sigmoid(numpy.dot(l0, synapseZero))
l2 = sigmoid(numpy.dot(l1, synapseOne))
l2_error = [trainY[i] - l2[i] for i in range(len(trainY))]
l2_delta = l2_error * sigmoid(l2, deriv=True)
l1_error = l2_delta.dot(synapseOne.T)
l1_delta = l1_error * sigmoid(l1, deriv=True)
synapseOne += l1.T.dot(l2_delta)
synapseZero += l0.T.dot(l1_delta)
synapse = numpy.dot(synapseZero, synapseOne)
a = sigmoid(numpy.dot(testX, synapse))
predictions = [
0 if abs(a[i]) < abs(a[i] - 1) else 1 for i in range(len(testX))
]
print(
"Categories: ", category1, "/", category2, " Error: ",
sum([
1 if predictions[i] != testY[i] else 0 for i in range(len(testY))
]), " out of ", len(testY), "accuracy", 100 - sum([
1 if predictions[i] != testY[i] else 0 for i in range(len(testY))
]) / len(testY) * 100, "%")
for key in data:
fig = pylab.figure()
values = data[key]
pylab.plot(values)
print key
count += 1
if count == 2:
break # print first example (Subject 4, Selective Attention, WORDS, Second Run)
pylab.show()
def twoDHeatMap(self, data, vmin=None, vmax=None, title=None, xLabel=None, yLabel=None, filename=None):
if not vmin is None and not vmax is None:
pylab.pcolor(numpy.array(data), vmin=vmin, vmax=vmax)
else:
pylab.pcolor(numpy.array(data))
if not title is None:
pylab.title(title)
if not xLabel is None:
pylab.xlabel(xLabel)
if not yLabel is None:
pylab.ylabel(yLabel)
if not filename is None:
pylab.savefig(filename + '.png')
pylab.show()
if __name__ == '__main__':
data = util.getVoxelArray()
data = util.dataSmoothing(data)
util.normalize(data)
g = Grapher()
g.twoDGraph(data)