naive_bayes.py 文件源码

def maximize_likelihood(self, data, responsibilities, weights, cmask=None):

        if not (cmask is None or cmask.shape == () or np.all(cmask)):  # cluster reduction
            responsibilities = responsibilities[:, cmask]
            self.names = list(compress(self.names, cmask))  # TODO: make self.names a numpy array?

        weights_combined = responsibilities * weights

        self.variables = np.dot(weights_combined.T, data.frequencies)
        with np.errstate(invalid='ignore'):  # if no training data is available for any class
            np.divide(self.variables, weights_combined.sum(axis=0, keepdims=True, dtype=types.large_float_type).T, out=self.variables)  # normalize before update, self.variables is types.prob_type

        dimchange = self.update()  # create cache for likelihood calculations

        # TODO: refactor this block
        ll = self.log_likelihood(data)
        std_per_class = common.weighted_std(ll, weights_combined)
        weight_per_class = weights_combined.sum(axis=0, dtype=types.large_float_type)
        weight_per_class /= weight_per_class.sum()
        std_per_class_mask = np.isnan(std_per_class)
        skipped_classes = std_per_class_mask.sum()
        self.stdev = np.ma.dot(np.ma.MaskedArray(std_per_class, mask=std_per_class_mask), weight_per_class)
        stderr.write("LOG %s: mean class likelihood standard deviation is %.2f (omitted %i/%i classes due to invalid or unsufficient data)\n" % (self._short_name, self.stdev, skipped_classes, self.num_components - skipped_classes))
        return dimchange, ll