def extractFeatures(self, tasks):
interactVec = Metric.interacMetric(tasks)
moments = self.computeMoments(tasks)
fourier = self.computeFourier(tasks)
norm = np.linalg.norm(moments)
moments = moments / (norm + 1e-15)
norm = np.linalg.norm(fourier)
fourier = fourier / (norm + 1e-15)
features = moments.tolist() + fourier.tolist()
# Idea to improve the feature vector
# if necessary
#nWeaklyInter = len([inter for inter in interacMetric if inter >= 1.0/4])
#nInter = len([inter for inter in interacMetric if inter >= 1.0/2])
norm = np.linalg.norm(features)
# normalize the features
# to make learning more robust
features = features / (norm + 1e-15)
self.features = features
return self.features
def computeMoments(self, tasks):
interactVec = Metric.interacMetric(tasks)
# use the counting measure
n = len(interactVec)
# note that we work with the count measure,
# thus the following moment computation is
# correct
for i in xrange(1,WorkloadFeatures.NumMoment+1):
self.moments[i-1] = sum([inter**i for inter \
in interactVec ])/float(n)
return self.moments
def computeFourier(self, tasks):
interactVec = Metric.interacMetric(tasks)
assert len(tasks) > 0
coeff = list(fft(interactVec))
# frequency zero corresponding to
# the integral of the original signal
coeff[0] = 0
coeff = np.abs(coeff)
maxFreq = np.argmax(coeff)
maxCoeff = np.max(coeff)
n = float(len(coeff))
self.fourier = [ maxFreq/n, maxCoeff ]
return self.fourier
