Beispiel #1
0
 def update_h(self):                    
     # assign samples to best matching centres ...
     m = 1.75
     tmp_dist = dist.pdist(self.W, self.data, metric='l2') + self._EPS
     self.H[:,:] = 0.0
     
     for i in range(self._num_bases):
         for k in range(self._num_bases):                
             self.H[i,:] += (tmp_dist[i,:]/tmp_dist[k,:])**(2.0/(m-1))
         
     self.H = np.where(self.H>0, 1.0/self.H, 0)    
Beispiel #2
0
 def _update_h(self):                    
     # assign samples to best matching centres ...
     m = 1.75
     tmp_dist = dist.pdist(self.W, self.data, metric='l2') + self._EPS
     self.H[:,:] = 0.0
     
     for i in range(self._num_bases):
         for k in range(self._num_bases):                
             self.H[i,:] += (tmp_dist[i,:]/tmp_dist[k,:])**(2.0/(m-1))
         
     self.H = np.where(self.H>0, 1.0/self.H, 0)    
Beispiel #3
0
    def update_w(self):
        
        def h(sel, D, k):
            # compute the volume for a selection of sel columns
            # and a k-1 simplex (-> k columns have to be selected)
            mv = np.max(D)
           
            # fill the remaining distance by the maximal overall found distance
            d = np.zeros((k,k)) + mv
            for i in range(k):
                d[i,i] = 0.0
            
            for idx_i,i in enumerate(sel):
                for idx_j,j in enumerate(sel):
                    d[idx_i,idx_j] = D[i, j]
    
            return d
  
        # compute distance matrix -> required for the volume
        D = pdist(self.data, self.data)
        Openset = {} 
        
        for i in range(self._num_samples):
            # compute volume for temp selection
            d = h([i],D,self._num_bases)           
            Openset[tuple([i])] = cmdet(d)
        
        Closedset = {}
        finished = False
        self._v = []
        self.init_sivm()
        next_sel = np.array([self.select[0]])
        niter = 0

        while not finished:
            # add the current selection to closedset
            Closedset[(tuple(next_sel))] = []

            for i in range(D.shape[0]):            
                # create a temp selection
                tmp_sel = np.array(next_sel).flatten()
                tmp_sel = np.concatenate((tmp_sel, [i]),axis=0)
                tmp_sel = np.unique(tmp_sel)
                tmp_sel = list(tmp_sel)
                hkey = tuple(tmp_sel)

                if len(tmp_sel) > len(next_sel) and (
                    not Closedset.has_key(hkey)) and (
                    not Openset.has_key(hkey)):
                    
                    # compute volume for temp selection
                    d = h(tmp_sel, D, self._num_bases)
      
                    # add to openset
                    Openset[hkey] = cmdet(d)

            # get next best tuple
            vmax = 0.0
            for (k,v) in Openset.iteritems():
                if v > vmax:
                    next_sel = k
                    vmax = v

            self._logger.info('Iter:' + str(niter))
            self._logger.info('Current selection:' + str(next_sel))
            self._logger.info('Current volume:' + str(vmax))
            self._v.append(vmax)

            # remove next_sel from openset
            Openset.pop(next_sel)

            if len(list(next_sel)) == self._num_bases:
                finished = True
            niter += 1

        # update some values ...
        self.select = list(next_sel)
        self.W = self.data[:, self.select] 
Beispiel #4
0
    def _update_w(self):
        def h(sel, D, k):
            # compute the volume for a selection of sel columns
            # and a k-1 simplex (-> k columns have to be selected)
            mv = np.max(D)

            # fill the remaining distance by the maximal overall found distance
            d = np.zeros((k, k)) + mv
            for i in range(k):
                d[i, i] = 0.0

            for idx_i, i in enumerate(sel):
                for idx_j, j in enumerate(sel):
                    d[idx_i, idx_j] = D[i, j]

            return d

        # compute distance matrix -> required for the volume
        D = pdist(self.data, self.data)
        Openset = {}

        for i in range(self._num_samples):
            # compute volume for temp selection
            d = h([i], D, self._num_bases)
            Openset[tuple([i])] = cmdet(d)

        Closedset = {}
        finished = False
        self._v = []
        self.init_sivm()
        next_sel = np.array([self.select[0]])
        niter = 0

        while not finished:
            # add the current selection to closedset
            Closedset[(tuple(next_sel))] = []

            for i in range(D.shape[0]):
                # create a temp selection
                tmp_sel = np.array(next_sel).flatten()
                tmp_sel = np.concatenate((tmp_sel, [i]), axis=0)
                tmp_sel = np.unique(tmp_sel)
                tmp_sel = list(tmp_sel)
                hkey = tuple(tmp_sel)

                if len(tmp_sel) > len(next_sel) and (
                        not Closedset.has_key(hkey)) and (
                            not Openset.has_key(hkey)):

                    # compute volume for temp selection
                    d = h(tmp_sel, D, self._num_bases)

                    # add to openset
                    Openset[hkey] = cmdet(d)

            # get next best tuple
            vmax = 0.0
            for (k, v) in Openset.iteritems():
                if v > vmax:
                    next_sel = k
                    vmax = v

            self._logger.info('Iter:' + str(niter))
            self._logger.info('Current selection:' + str(next_sel))
            self._logger.info('Current volume:' + str(vmax))
            self._v.append(vmax)

            # remove next_sel from openset
            Openset.pop(next_sel)

            if len(list(next_sel)) == self._num_bases:
                finished = True
            niter += 1

        # update some values ...
        self.select = list(next_sel)
        self.W = self.data[:, self.select]