使用sklearn.metrics.pairwise_distance或scipy.spatial.distance.cdist估计大型daraset的成对距离

Question

我试图估计一个约300000张图像的数据集的特征之间的成对距离，以计算它们的最小值。但是，计算的距离矩阵太大，无法放入内存，导致OOM错误

在上下文中，我试图实现kcenter_贪婪方法来进行主动学习，如这里（https://github.com/google/active-learning/blob/master/sampling_methods/kcenter_greedy.py）

人们评估过的大多数学术数据集都很小，因此代码运行良好。在我的例子中，我希望使用一个更大的数据集，对于这个数据集，sklearn.metrics.pairwise_distances函数没有那么有用

下面是代码的相关部分

def update_distances(self, cluster_centers, only_new=True, reset_dist=False):
    """Update min distances given cluster centers.
    Args:
      cluster_centers: indices of cluster centers
      only_new: only calculate distance for newly selected points and update
        min_distances.
      rest_dist: whether to reset min_distances.
    """

    if reset_dist:
      self.min_distances = None
    if only_new:
      cluster_centers = [d for d in cluster_centers
                         if d not in self.already_selected]
    if cluster_centers:
      # Update min_distances for all examples given new cluster center.
      x = self.features[cluster_centers]
      ################ This is the section causing OOM issues######################################
      dist = pairwise_distances(self.features, x, metric=self.metric)

      if self.min_distances is None:
        self.min_distances = np.min(dist, axis=1).reshape(-1,1)
      else:
        self.min_distances = np.minimum(self.min_distances, dist)

  def select_batch_(self, model, already_selected, N, **kwargs):
    """
    Diversity promoting active learning method that greedily forms a batch
    to minimize the maximum distance to a cluster center among all unlabeled
    datapoints.
    Args:
      model: model with scikit-like API with decision_function implemented
      already_selected: index of datapoints already selected
      N: batch size
    Returns:
      indices of points selected to minimize distance to cluster centers
    """

    try:
      # Assumes that the transform function takes in original data and not
      # flattened data.
      print('Getting transformed features...')
      self.features = model.transform(self.X)
      print('Calculating distances...')
      self.update_distances(already_selected, only_new=False, reset_dist=True)
    except:
      print('Using flat_X as features.')
      self.update_distances(already_selected, only_new=True, reset_dist=False)

    new_batch = []

    for _ in range(N):
      if self.already_selected is None:
        # Initialize centers with a randomly selected datapoint
        ind = np.random.choice(np.arange(self.n_obs))
      else:
        ind = np.argmax(self.min_distances)
      # New examples should not be in already selected since those points
      # should have min_distance of zero to a cluster center.
      assert ind not in already_selected

      self.update_distances([ind], only_new=True, reset_dist=False)
      new_batch.append(ind)
    print('Maximum distance from cluster centers is %0.2f'
            % max(self.min_distances))


    self.already_selected = already_selected

    return new_batch

我也尝试过使用scipy.spatial.distance.cdist函数，但这无助于解决OOM问题

有没有更好的方法更有效地找到wrt内存的最小距离？过去scikit-learn库中的优化对我有所帮助，但在这种情况下，它似乎不适用于大型数据集