Python sklearn 模块，decomposition() 实例源码

def load(self, modelfile):
      with open(modelfile, 'rb') as pklfile:
          self.model = pickle.load(pklfile)
      assert( isinstance(self.model,sklearn.decomposition.pca.PCA) )
      print "Successfully loaded pca model from %s." % modelfile

def _auto_low_rank_model(data, mode, n_jobs, method_params, cv,
                         stop_early=True, verbose=None):
    """compute latent variable models."""
    method_params = cp.deepcopy(method_params)
    iter_n_components = method_params.pop('iter_n_components')
    if iter_n_components is None:
        iter_n_components = np.arange(5, data.shape[1], 5)
    from sklearn.decomposition import PCA, FactorAnalysis
    if mode == 'factor_analysis':
        est = FactorAnalysis
    elif mode == 'pca':
        est = PCA
    else:
        raise ValueError('Come on, this is not a low rank estimator: %s' %
                         mode)
    est = est(**method_params)
    est.n_components = 1
    scores = np.empty_like(iter_n_components, dtype=np.float64)
    scores.fill(np.nan)

    # make sure we don't empty the thing if it's a generator
    max_n = max(list(cp.deepcopy(iter_n_components)))
    if max_n > data.shape[1]:
        warn('You are trying to estimate %i components on matrix '
             'with %i features.' % (max_n, data.shape[1]))

    for ii, n in enumerate(iter_n_components):
        est.n_components = n
        try:  # this may fail depending on rank and split
            score = _cross_val(data=data, est=est, cv=cv, n_jobs=n_jobs)
        except ValueError:
            score = np.inf
        if np.isinf(score) or score > 0:
            logger.info('... infinite values encountered. stopping estimation')
            break
        logger.info('... rank: %i - loglik: %0.3f' % (n, score))
        if score != -np.inf:
            scores[ii] = score

        if (ii >= 3 and np.all(np.diff(scores[ii - 3:ii]) < 0.) and
           stop_early is True):
            # early stop search when loglik has been going down 3 times
            logger.info('early stopping parameter search.')
            break

    # happens if rank is too low right form the beginning
    if np.isnan(scores).all():
        raise RuntimeError('Oh no! Could not estimate covariance because all '
                           'scores were NaN. Please contact the MNE-Python '
                           'developers.')

    i_score = np.nanargmax(scores)
    best = est.n_components = iter_n_components[i_score]
    logger.info('... best model at rank = %i' % best)
    runtime_info = {'ranks': np.array(iter_n_components),
                    'scores': scores,
                    'best': best,
                    'cv': cv}
    return est, runtime_info