models.py 文件源码

def compute_super_codebook(self, feature_size):
        '''
        Merges the GMMs that were computed for each class
        into one GMM (super codebook).
        @param feature_size: dimensionality of a feature vector 
        '''

        # Get number of classes
        ncl = len(self.labels)
        # Compute vocabsize per class
        vocab_size_per_cl = max(1, self.vocab_size / ncl)
        # Create GMM for overall repr
        print >> sys.stderr, 'Using GMM with', self.vocab_size, 'and', vocab_size_per_cl , 'per class.'
        self.gmm = GMM(self.vocab_size, n_iter=1, params='w', covariance_type='diag')
        # Init by fitting with ones
        self.gmm.fit(np.ones((self.vocab_size, feature_size)))

        # Overwrite values from supervised codebook GMM by class GMMs
        index = 0
        for _, sgmm in self.gmms.iteritems():
            vocab_size_per_cl = len(sgmm.means_)
            self.gmm.means_  [index:index + vocab_size_per_cl, :] = sgmm.means_
            self.gmm.covars_ [index:index + vocab_size_per_cl, :] = sgmm.covars_
            index += vocab_size_per_cl

        # Set uniform GMM weights
        self.gmm.weights_ = np.ones(self.vocab_size) / float(self.vocab_size)
        return