[fix] densite de proba somme a 1

Author: Pierre Gerard

src/densite.py

@@ -23,31 +23,32 @@ def _getOneDimDate(self):
 
 
  def getOneDimGraph(self):
- trainSetPlot, = plt.plot(self.oneDimData,numpy.array([i/500. for i in range(len(self.oneDimData))]), 'bo')
+ trainSetPlot, = plt.plot(self.oneDimData,numpy.array([i/1000. for i in range(len(self.oneDimData))]), 'bo')
 
  dg = densite_fonction.DensiteGaussienne(1)
  dg.train(self.oneDimData)
+ print(dg.getSigma())
  x = numpy.linspace(3,7,1000)
  densParamPlot, = plt.plot(x,[dg.p(i) for i in x], 'red', label= "Densite parametrique")
 
 
- sigma = 0.2
+ sigma = 0.05
  dg = densite_fonction.DensiteParzen(1,sigma)
  dg.train(self.oneDimData)
  x = numpy.linspace(3,7,1000)
- parzenParamPlotGrand, = plt.plot(x,[dg.p(i) for i in x], 'blue', label= "Parzen sigma petit")
+ parzenParamPlotGrand, = plt.plot(x,[dg.p(i) for i in x], 'yellow', label= "Parzen sigma petit")
 
  sigma = 1
  dg = densite_fonction.DensiteParzen(1,sigma)
  dg.train(self.oneDimData)
  x = numpy.linspace(3,7,1000)
  parzenParamPlotPetit, = plt.plot(x,[dg.p(i) for i in x], 'green', label= "Parzen sigma grand")
 
- sigma = 0.45
+ sigma = 0.349
  dg = densite_fonction.DensiteParzen(1,sigma)
  dg.train(self.oneDimData)
  x = numpy.linspace(3,7,1000)
- parzenParamPlotAdequat, = plt.plot(x,[dg.p(i) for i in x], 'yellow', label= "Parzen sigma adequat")
+ parzenParamPlotAdequat, = plt.plot(x,[dg.p(i) for i in x], 'blue', label= "Parzen sigma adequat")
 
  plt.legend([trainSetPlot, densParamPlot, parzenParamPlotGrand,parzenParamPlotPetit,parzenParamPlotAdequat], ["Ensemble entrainement", "Densite parametrique","Parzen sigma petit","Parzen sigma grand","Parzen sigma adequat"])
  plt.axis([3, 7, 0, 1.7])

src/densite_fonction.py

@@ -24,8 +24,12 @@ def getSigma(self):
 
  def p(self, x):
  sigma = numpy.sqrt(self._sigma_sq)
- # TODO dimension d
- return 1./(numpy.sqrt(2.*numpy.pi)*sigma)* numpy.exp(-numpy.power((x - self._mu)/sigma, 2.)/2)
+ if self._n_dims == 1:
+ expo = -(float(x) - self._mu)**2
+ expo /= 2*(float(sigma)**2)
+ return 1./(numpy.sqrt(2.*numpy.pi)*sigma)* numpy.exp(expo)
+ else:
+ pass # TODO dimension d
 
  def compute_predictions(self, test_data):
  """
@@ -52,10 +56,15 @@ def train(self, train_data):
  def p(self, x):
  res = 0
  for i,xi in enumerate(self._data):
- res += 1./((2.*numpy.pi)**(self._n_dims/2.) * (self._sigma**self._n_dims)) \
- * numpy.exp(-numpy.power(self._distanceFunction(x, xi)/self._sigma**2, 2.)/2)
+ expo = -1/2. * self._distanceFunction(x, xi)**2
+ expo /= (self._sigma**2)
+ res += 1./(((2.*numpy.pi)**(self._n_dims/2.)) * (self._sigma**self._n_dims)) \
+ * numpy.exp(expo)
  res /= float(len(self._data))
  return res
 
  def compute_predictions(self, test_data):
- pass
+ log_prob = []
+ for i in test_data:
+ log_prob.append(numpy.log(i))
+ return res