标签传播算法(Label Propagation)及Python实现

转:https://blog.csdn.net/zouxy09/article/details/49105265#commentBox

标签传播算法(Label Propagation)及Python实现

众所周知,机器学习可以大体分为三大类:监督学习、非监督学习和半监督学习。监督学习可以认为是我们有非常多的labeled标注数据来train一个模型,期待这个模型能学习到数据的分布,以期对未来没有见到的样本做预测。那这个性能的源头--训练数据,就显得非常感觉。你必须有足够的训练数据,以覆盖真正现实数据中的样本分布才可以,这样学习到的模型才有意义。那非监督学习就是没有任何的labeled数据,就是平时所说的聚类了,利用他们本身的数据分布,给他们划分类别。而半监督学习,顾名思义就是处于两者之间的,只有少量的labeled数据,我们试图从这少量的labeled数据和大量的unlabeled数据中学习到有用的信息。

一、半监督学习

       半监督学习(Semi-supervised learning)发挥作用的场合是:你的数据有一些有label,一些没有。而且一般是绝大部分都没有,只有少许几个有label。半监督学习算法会充分的利用unlabeled数据来捕捉我们整个数据的潜在分布。它基于三大假设:

       1)Smoothness平滑假设:相似的数据具有相同的label。

       2)Cluster聚类假设:处于同一个聚类下的数据具有相同label。

       3)Manifold流形假设:处于同一流形结构下的数据具有相同label。

       例如下图,只有两个labeled数据,如果直接用他们来训练一个分类器,例如LR或者SVM,那么学出来的分类面就是左图那样的。如果现实中,这个数据是右图那边分布的话,猪都看得出来,左图训练的这个分类器烂的一塌糊涂、惨不忍睹。因为我们的labeled训练数据太少了,都没办法覆盖我们未来可能遇到的情况。但是,如果右图那样,把大量的unlabeled数据(黑色的)都考虑进来,有个全局观念,牛逼的算法会发现,哎哟,原来是两个圈圈(分别处于两个圆形的流形之上)!那算法就很聪明,把大圈的数据都归类为红色类别,把内圈的数据都归类为蓝色类别。因为,实践中,labeled数据是昂贵,很难获得的,但unlabeled数据就不是了,写个脚本在网上爬就可以了,因此如果能充分利用大量的unlabeled数据来辅助提升我们的模型学习,这个价值就非常大。

标签传播算法(Label Propagation)及Python实现

       半监督学习算法有很多,下面我们介绍最简单的标签传播算法(label propagation),最喜欢简单了,哈哈。

二、标签传播算法

       标签传播算法(label propagation)的核心思想非常简单:相似的数据应该具有相同的label。LP算法包括两大步骤:1)构造相似矩阵;2)勇敢的传播吧。

2.1、相似矩阵构建

       LP算法是基于Graph的,因此我们需要先构建一个图。我们为所有的数据构建一个图,图的节点就是一个数据点,包含labeled和unlabeled的数据。节点i和节点j的边表示他们的相似度。这个图的构建方法有很多,这里我们假设这个图是全连接的,节点i和节点j的边权重为:

标签传播算法(Label Propagation)及Python实现

       这里,α是超参。

       还有个非常常用的图构建方法是knn图,也就是只保留每个节点的k近邻权重,其他的为0,也就是不存在边,因此是稀疏的相似矩阵。

2.2、LP算法

       标签传播算法非常简单:通过节点之间的边传播label。边的权重越大,表示两个节点越相似,那么label越容易传播过去。我们定义一个NxN的概率转移矩阵P:

标签传播算法(Label Propagation)及Python实现

       Pij表示从节点i转移到节点j的概率。假设有C个类和L个labeled样本,我们定义一个LxC的label矩阵YL,第i行表示第i个样本的标签指示向量,即如果第i个样本的类别是j,那么该行的第j个元素为1,其他为0。同样,我们也给U个unlabeled样本一个UxC的label矩阵YU。把他们合并,我们得到一个NxC的soft label矩阵F=[YL;YU]。soft label的意思是,我们保留样本i属于每个类别的概率,而不是互斥性的,这个样本以概率1只属于一个类。当然了,最后确定这个样本i的类别的时候,是取max也就是概率最大的那个类作为它的类别的。那F里面有个YU,它一开始是不知道的,那最开始的值是多少?无所谓,随便设置一个值就可以了。

       千呼万唤始出来,简单的LP算法如下:

       1)执行传播:F=PF

       2)重置F中labeled样本的标签:FL=YL

       3)重复步骤1)和2)直到F收敛。

       步骤1)就是将矩阵P和矩阵F相乘,这一步,每个节点都将自己的label以P确定的概率传播给其他节点。如果两个节点越相似(在欧式空间中距离越近),那么对方的label就越容易被自己的label赋予,就是更容易拉帮结派。步骤2)非常关键,因为labeled数据的label是事先确定的,它不能被带跑,所以每次传播完,它都得回归它本来的label。随着labeled数据不断的将自己的label传播出去,最后的类边界会穿越高密度区域,而停留在低密度的间隔中。相当于每个不同类别的labeled样本划分了*范围。

2.3、变身的LP算法

       我们知道,我们每次迭代都是计算一个soft label矩阵F=[YL;YU],但是YL是已知的,计算它没有什么用,在步骤2)的时候,还得把它弄回来。我们关心的只是YU,那我们能不能只计算YU呢?Yes。我们将矩阵P做以下划分:

标签传播算法(Label Propagation)及Python实现

       这时候,我们的算法就一个运算:

标签传播算法(Label Propagation)及Python实现

       迭代上面这个步骤直到收敛就ok了,是不是很cool。可以看到FU不但取决于labeled数据的标签及其转移概率,还取决了unlabeled数据的当前label和转移概率。因此LP算法能额外运用unlabeled数据的分布特点。

       这个算法的收敛性也非常容易证明,具体见参考文献[1]。实际上,它是可以收敛到一个凸解的:

标签传播算法(Label Propagation)及Python实现

       所以我们也可以直接这样求解,以获得最终的YU。但是在实际的应用过程中,由于矩阵求逆需要O(n3)的复杂度,所以如果unlabeled数据非常多,那么I – PUU矩阵的求逆将会非常耗时,因此这时候一般选择迭代算法来实现。

三、LP算法的Python实现

       Python环境的搭建就不啰嗦了,可以参考前面的博客。需要额外依赖的库是经典的numpy和matplotlib。代码中包含了两种图的构建方法:RBF和KNN指定。同时,自己生成了两个toy数据库:两条长形形状和两个圈圈的数据。第四部分我们用大点的数据库来做实验,先简单的可视化验证代码的正确性,再前线。

       算法代码:

  1. #***************************************************************************
  2. #*
  3. #* Description: label propagation
  4. #* Author: Zou Xiaoyi ([email protected])
  5. #* Date: 2015-10-15
  6. #* HomePage: http://blog.csdn.net/zouxy09
  7. #*
  8. #**************************************************************************
  9. import time
  10. import numpy as np
  11. # return k neighbors index
  12. def navie_knn(dataSet, query, k):
  13. numSamples = dataSet.shape[0]
  14. ## step 1: calculate Euclidean distance
  15. diff = np.tile(query, (numSamples, 1)) - dataSet
  16. squaredDiff = diff ** 2
  17. squaredDist = np.sum(squaredDiff, axis = 1) # sum is performed by row
  18. ## step 2: sort the distance
  19. sortedDistIndices = np.argsort(squaredDist)
  20. if k > len(sortedDistIndices):
  21. k = len(sortedDistIndices)
  22. return sortedDistIndices[0:k]
  23. # build a big graph (normalized weight matrix)
  24. def buildGraph(MatX, kernel_type, rbf_sigma = None, knn_num_neighbors = None):
  25. num_samples = MatX.shape[0]
  26. affinity_matrix = np.zeros((num_samples, num_samples), np.float32)
  27. if kernel_type == 'rbf':
  28. if rbf_sigma == None:
  29. raise ValueError('You should input a sigma of rbf kernel!')
  30. for i in xrange(num_samples):
  31. row_sum = 0.0
  32. for j in xrange(num_samples):
  33. diff = MatX[i, :] - MatX[j, :]
  34. affinity_matrix[i][j] = np.exp(sum(diff**2) / (-2.0 * rbf_sigma**2))
  35. row_sum += affinity_matrix[i][j]
  36. affinity_matrix[i][:] /= row_sum
  37. elif kernel_type == 'knn':
  38. if knn_num_neighbors == None:
  39. raise ValueError('You should input a k of knn kernel!')
  40. for i in xrange(num_samples):
  41. k_neighbors = navie_knn(MatX, MatX[i, :], knn_num_neighbors)
  42. affinity_matrix[i][k_neighbors] = 1.0 / knn_num_neighbors
  43. else:
  44. raise NameError('Not support kernel type! You can use knn or rbf!')
  45. return affinity_matrix
  46. # label propagation
  47. def labelPropagation(Mat_Label, Mat_Unlabel, labels, kernel_type = 'rbf', rbf_sigma = 1.5, \
  48. knn_num_neighbors = 10, max_iter = 500, tol = 1e-3):
  49. # initialize
  50. num_label_samples = Mat_Label.shape[0]
  51. num_unlabel_samples = Mat_Unlabel.shape[0]
  52. num_samples = num_label_samples + num_unlabel_samples
  53. labels_list = np.unique(labels)
  54. num_classes = len(labels_list)
  55. MatX = np.vstack((Mat_Label, Mat_Unlabel))
  56. clamp_data_label = np.zeros((num_label_samples, num_classes), np.float32)
  57. for i in xrange(num_label_samples):
  58. clamp_data_label[i][labels[i]] = 1.0
  59. label_function = np.zeros((num_samples, num_classes), np.float32)
  60. label_function[0 : num_label_samples] = clamp_data_label
  61. label_function[num_label_samples : num_samples] = -1
  62. # graph construction
  63. affinity_matrix = buildGraph(MatX, kernel_type, rbf_sigma, knn_num_neighbors)
  64. # start to propagation
  65. iter = 0; pre_label_function = np.zeros((num_samples, num_classes), np.float32)
  66. changed = np.abs(pre_label_function - label_function).sum()
  67. while iter < max_iter and changed > tol:
  68. if iter % 1 == 0:
  69. print "---> Iteration %d/%d, changed: %f" % (iter, max_iter, changed)
  70. pre_label_function = label_function
  71. iter += 1
  72. # propagation
  73. label_function = np.dot(affinity_matrix, label_function)
  74. # clamp
  75. label_function[0 : num_label_samples] = clamp_data_label
  76. # check converge
  77. changed = np.abs(pre_label_function - label_function).sum()
  78. # get terminate label of unlabeled data
  79. unlabel_data_labels = np.zeros(num_unlabel_samples)
  80. for i in xrange(num_unlabel_samples):
  81. unlabel_data_labels[i] = np.argmax(label_function[i+num_label_samples])
  82. return unlabel_data_labels

       测试代码:

  1. #***************************************************************************
  2. #*
  3. #* Description: label propagation
  4. #* Author: Zou Xiaoyi ([email protected])
  5. #* Date: 2015-10-15
  6. #* HomePage: http://blog.csdn.net/zouxy09
  7. #*
  8. #**************************************************************************
  9. import time
  10. import math
  11. import numpy as np
  12. from label_propagation import labelPropagation
  13. # show
  14. def show(Mat_Label, labels, Mat_Unlabel, unlabel_data_labels):
  15. import matplotlib.pyplot as plt
  16. for i in range(Mat_Label.shape[0]):
  17. if int(labels[i]) == 0:
  18. plt.plot(Mat_Label[i, 0], Mat_Label[i, 1], 'Dr')
  19. elif int(labels[i]) == 1:
  20. plt.plot(Mat_Label[i, 0], Mat_Label[i, 1], 'Db')
  21. else:
  22. plt.plot(Mat_Label[i, 0], Mat_Label[i, 1], 'Dy')
  23. for i in range(Mat_Unlabel.shape[0]):
  24. if int(unlabel_data_labels[i]) == 0:
  25. plt.plot(Mat_Unlabel[i, 0], Mat_Unlabel[i, 1], 'or')
  26. elif int(unlabel_data_labels[i]) == 1:
  27. plt.plot(Mat_Unlabel[i, 0], Mat_Unlabel[i, 1], 'ob')
  28. else:
  29. plt.plot(Mat_Unlabel[i, 0], Mat_Unlabel[i, 1], 'oy')
  30. plt.xlabel('X1'); plt.ylabel('X2')
  31. plt.xlim(0.0, 12.)
  32. plt.ylim(0.0, 12.)
  33. plt.show()
  34. def loadCircleData(num_data):
  35. center = np.array([5.0, 5.0])
  36. radiu_inner = 2
  37. radiu_outer = 4
  38. num_inner = num_data / 3
  39. num_outer = num_data - num_inner
  40. data = []
  41. theta = 0.0
  42. for i in range(num_inner):
  43. pho = (theta % 360) * math.pi / 180
  44. tmp = np.zeros(2, np.float32)
  45. tmp[0] = radiu_inner * math.cos(pho) + np.random.rand(1) + center[0]
  46. tmp[1] = radiu_inner * math.sin(pho) + np.random.rand(1) + center[1]
  47. data.append(tmp)
  48. theta += 2
  49. theta = 0.0
  50. for i in range(num_outer):
  51. pho = (theta % 360) * math.pi / 180
  52. tmp = np.zeros(2, np.float32)
  53. tmp[0] = radiu_outer * math.cos(pho) + np.random.rand(1) + center[0]
  54. tmp[1] = radiu_outer * math.sin(pho) + np.random.rand(1) + center[1]
  55. data.append(tmp)
  56. theta += 1
  57. Mat_Label = np.zeros((2, 2), np.float32)
  58. Mat_Label[0] = center + np.array([-radiu_inner + 0.5, 0])
  59. Mat_Label[1] = center + np.array([-radiu_outer + 0.5, 0])
  60. labels = [0, 1]
  61. Mat_Unlabel = np.vstack(data)
  62. return Mat_Label, labels, Mat_Unlabel
  63. def loadBandData(num_unlabel_samples):
  64. #Mat_Label = np.array([[5.0, 2.], [5.0, 8.0]])
  65. #labels = [0, 1]
  66. #Mat_Unlabel = np.array([[5.1, 2.], [5.0, 8.1]])
  67. Mat_Label = np.array([[5.0, 2.], [5.0, 8.0]])
  68. labels = [0, 1]
  69. num_dim = Mat_Label.shape[1]
  70. Mat_Unlabel = np.zeros((num_unlabel_samples, num_dim), np.float32)
  71. Mat_Unlabel[:num_unlabel_samples/2, :] = (np.random.rand(num_unlabel_samples/2, num_dim) - 0.5) * np.array([3, 1]) + Mat_Label[0]
  72. Mat_Unlabel[num_unlabel_samples/2 : num_unlabel_samples, :] = (np.random.rand(num_unlabel_samples/2, num_dim) - 0.5) * np.array([3, 1]) + Mat_Label[1]
  73. return Mat_Label, labels, Mat_Unlabel
  74. # main function
  75. if __name__ == "__main__":
  76. num_unlabel_samples = 800
  77. #Mat_Label, labels, Mat_Unlabel = loadBandData(num_unlabel_samples)
  78. Mat_Label, labels, Mat_Unlabel = loadCircleData(num_unlabel_samples)
  79. ## Notice: when use 'rbf' as our kernel, the choice of hyper parameter 'sigma' is very import! It should be
  80. ## chose according to your dataset, specific the distance of two data points. I think it should ensure that
  81. ## each point has about 10 knn or w_i,j is large enough. It also influence the speed of converge. So, may be
  82. ## 'knn' kernel is better!
  83. #unlabel_data_labels = labelPropagation(Mat_Label, Mat_Unlabel, labels, kernel_type = 'rbf', rbf_sigma = 0.2)
  84. unlabel_data_labels = labelPropagation(Mat_Label, Mat_Unlabel, labels, kernel_type = 'knn', knn_num_neighbors = 10, max_iter = 400)
  85. show(Mat_Label, labels, Mat_Unlabel, unlabel_data_labels)

       该注释的,代码都注释的,有看不明白的,欢迎交流。不同迭代次数时候的结果如下:

标签传播算法(Label Propagation)及Python实现

       是不是很漂亮的传播过程?!在数值上也是可以看到随着迭代的进行逐渐收敛的,迭代的数值变化过程如下:

  1. ---> Iteration 0/400, changed: 1602.000000
  2. ---> Iteration 1/400, changed: 6.300182
  3. ---> Iteration 2/400, changed: 5.129996
  4. ---> Iteration 3/400, changed: 4.301994
  5. ---> Iteration 4/400, changed: 3.819295
  6. ---> Iteration 5/400, changed: 3.501743
  7. ---> Iteration 6/400, changed: 3.277122
  8. ---> Iteration 7/400, changed: 3.105952
  9. ---> Iteration 8/400, changed: 2.967030
  10. ---> Iteration 9/400, changed: 2.848606
  11. ---> Iteration 10/400, changed: 2.743997
  12. ---> Iteration 11/400, changed: 2.649270
  13. ---> Iteration 12/400, changed: 2.562057
  14. ---> Iteration 13/400, changed: 2.480885
  15. ---> Iteration 14/400, changed: 2.404774
  16. ---> Iteration 15/400, changed: 2.333075
  17. ---> Iteration 16/400, changed: 2.265301
  18. ---> Iteration 17/400, changed: 2.201107
  19. ---> Iteration 18/400, changed: 2.140209
  20. ---> Iteration 19/400, changed: 2.082354
  21. ---> Iteration 20/400, changed: 2.027376
  22. ---> Iteration 21/400, changed: 1.975071
  23. ---> Iteration 22/400, changed: 1.925286
  24. ---> Iteration 23/400, changed: 1.877894
  25. ---> Iteration 24/400, changed: 1.832743
  26. ---> Iteration 25/400, changed: 1.789721
  27. ---> Iteration 26/400, changed: 1.748706
  28. ---> Iteration 27/400, changed: 1.709593
  29. ---> Iteration 28/400, changed: 1.672284
  30. ---> Iteration 29/400, changed: 1.636668
  31. ---> Iteration 30/400, changed: 1.602668
  32. ---> Iteration 31/400, changed: 1.570200
  33. ---> Iteration 32/400, changed: 1.539179
  34. ---> Iteration 33/400, changed: 1.509530
  35. ---> Iteration 34/400, changed: 1.481182
  36. ---> Iteration 35/400, changed: 1.454066
  37. ---> Iteration 36/400, changed: 1.428120
  38. ---> Iteration 37/400, changed: 1.403283
  39. ---> Iteration 38/400, changed: 1.379502
  40. ---> Iteration 39/400, changed: 1.356734
  41. ---> Iteration 40/400, changed: 1.334906
  42. ---> Iteration 41/400, changed: 1.313983
  43. ---> Iteration 42/400, changed: 1.293921
  44. ---> Iteration 43/400, changed: 1.274681
  45. ---> Iteration 44/400, changed: 1.256214
  46. ---> Iteration 45/400, changed: 1.238491
  47. ---> Iteration 46/400, changed: 1.221474
  48. ---> Iteration 47/400, changed: 1.205126
  49. ---> Iteration 48/400, changed: 1.189417
  50. ---> Iteration 49/400, changed: 1.174316
  51. ---> Iteration 50/400, changed: 1.159804
  52. ---> Iteration 51/400, changed: 1.145844
  53. ---> Iteration 52/400, changed: 1.132414
  54. ---> Iteration 53/400, changed: 1.119490
  55. ---> Iteration 54/400, changed: 1.107032
  56. ---> Iteration 55/400, changed: 1.095054
  57. ---> Iteration 56/400, changed: 1.083513
  58. ---> Iteration 57/400, changed: 1.072397
  59. ---> Iteration 58/400, changed: 1.061671
  60. ---> Iteration 59/400, changed: 1.051324
  61. ---> Iteration 60/400, changed: 1.041363
  62. ---> Iteration 61/400, changed: 1.031742
  63. ---> Iteration 62/400, changed: 1.022459
  64. ---> Iteration 63/400, changed: 1.013494
  65. ---> Iteration 64/400, changed: 1.004836
  66. ---> Iteration 65/400, changed: 0.996484
  67. ---> Iteration 66/400, changed: 0.988407
  68. ---> Iteration 67/400, changed: 0.980592
  69. ---> Iteration 68/400, changed: 0.973045
  70. ---> Iteration 69/400, changed: 0.965744
  71. ---> Iteration 70/400, changed: 0.958682
  72. ---> Iteration 71/400, changed: 0.951848
  73. ---> Iteration 72/400, changed: 0.945227
  74. ---> Iteration 73/400, changed: 0.938820
  75. ---> Iteration 74/400, changed: 0.932608
  76. ---> Iteration 75/400, changed: 0.926590
  77. ---> Iteration 76/400, changed: 0.920765
  78. ---> Iteration 77/400, changed: 0.915107
  79. ---> Iteration 78/400, changed: 0.909628
  80. ---> Iteration 79/400, changed: 0.904309
  81. ---> Iteration 80/400, changed: 0.899143
  82. ---> Iteration 81/400, changed: 0.894122
  83. ---> Iteration 82/400, changed: 0.889259
  84. ---> Iteration 83/400, changed: 0.884530
  85. ---> Iteration 84/400, changed: 0.879933
  86. ---> Iteration 85/400, changed: 0.875464
  87. ---> Iteration 86/400, changed: 0.871121
  88. ---> Iteration 87/400, changed: 0.866888
  89. ---> Iteration 88/400, changed: 0.862773
  90. ---> Iteration 89/400, changed: 0.858783
  91. ---> Iteration 90/400, changed: 0.854879
  92. ---> Iteration 91/400, changed: 0.851084
  93. ---> Iteration 92/400, changed: 0.847382
  94. ---> Iteration 93/400, changed: 0.843779
  95. ---> Iteration 94/400, changed: 0.840274
  96. ---> Iteration 95/400, changed: 0.836842
  97. ---> Iteration 96/400, changed: 0.833501
  98. ---> Iteration 97/400, changed: 0.830240
  99. ---> Iteration 98/400, changed: 0.827051
  100. ---> Iteration 99/400, changed: 0.823950
  101. ---> Iteration 100/400, changed: 0.820906
  102. ---> Iteration 101/400, changed: 0.817946
  103. ---> Iteration 102/400, changed: 0.815053
  104. ---> Iteration 103/400, changed: 0.812217
  105. ---> Iteration 104/400, changed: 0.809437
  106. ---> Iteration 105/400, changed: 0.806724
  107. ---> Iteration 106/400, changed: 0.804076
  108. ---> Iteration 107/400, changed: 0.801480
  109. ---> Iteration 108/400, changed: 0.798937
  110. ---> Iteration 109/400, changed: 0.796448
  111. ---> Iteration 110/400, changed: 0.794008
  112. ---> Iteration 111/400, changed: 0.791612
  113. ---> Iteration 112/400, changed: 0.789282
  114. ---> Iteration 113/400, changed: 0.786984
  115. ---> Iteration 114/400, changed: 0.784728
  116. ---> Iteration 115/400, changed: 0.782516
  117. ---> Iteration 116/400, changed: 0.780355
  118. ---> Iteration 117/400, changed: 0.778216
  119. ---> Iteration 118/400, changed: 0.776139
  120. ---> Iteration 119/400, changed: 0.774087
  121. ---> Iteration 120/400, changed: 0.772072
  122. ---> Iteration 121/400, changed: 0.770085
  123. ---> Iteration 122/400, changed: 0.768146
  124. ---> Iteration 123/400, changed: 0.766232
  125. ---> Iteration 124/400, changed: 0.764356
  126. ---> Iteration 125/400, changed: 0.762504
  127. ---> Iteration 126/400, changed: 0.760685
  128. ---> Iteration 127/400, changed: 0.758889
  129. ---> Iteration 128/400, changed: 0.757135
  130. ---> Iteration 129/400, changed: 0.755406

四、LP算法MPI并行实现

       这里,我们测试的是LP的变身版本。从公式,我们可以看到,第二项PULYL迭代过程并没有发生变化,所以这部分实际上从迭代开始就可以计算好,从而避免重复计算。不过,不管怎样,LP算法都要计算一个UxU的矩阵PUU和一个UxC矩阵FU的乘积。当我们的unlabeled数据非常多,而且类别也很多的时候,计算是很慢的,同时占用的内存量也非常大。另外,构造Graph需要计算两两的相似度,也是O(n2)的复杂度,当我们数据的特征维度很大的时候,这个计算量也是非常客观的。所以我们就得考虑并行处理了。而且最好是能放到集群上并行。那如何并行呢?

       对算法的并行化,一般分为两种:数据并行和模型并行。

       数据并行很好理解,就是将数据划分,每个节点只处理一部分数据,例如我们构造图的时候,计算每个数据的k近邻。例如我们有1000个样本和20个CPU节点,那么就平均分发,让每个CPU节点计算50个样本的k近邻,然后最后再合并大家的结果。可见这个加速比也是非常可观的。

       模型并行一般发生在模型很大,无法放到单机的内存里面的时候。例如庞大的深度神经网络训练的时候,就需要把这个网络切开,然后分别求解梯度,最后有个leader的节点来收集大家的梯度,再反馈给大家去更新。当然了,其中存在更细致和高效的工程处理方法。在我们的LP算法中,也是可以做模型并行的。假如我们的类别数C很大,把类别数切开,让不同的CPU节点处理,实际上就相当于模型并行了。

       那为啥不切大矩阵PUU,而是切小点的矩阵FU,因为大矩阵PUU没法独立分块,并行的一个原则是处理必须是独立的。 矩阵FU依赖的是所有的U,而把PUU切开分发到其他节点的时候,每次FU的更新都需要和其他的节点通信,这个通信的代价是很大的(实际上,很多并行系统没法达到线性的加速度的瓶颈是通信!线性加速比是,我增加了n台机器,速度就提升了n倍)。但是对类别C也就是矩阵FU切分,就不会有这个问题,因为他们的计算是独立的。只是决定样本的最终类别的时候,将所有的FU收集回来求max就可以了。

       所以,在下面的代码中,是同时包含了数据并行和模型并行的雏形的。另外,还值得一提的是,我们是迭代算法,那决定什么时候迭代算法停止?除了判断收敛外,我们还可以让每迭代几步,就用测试label测试一次结果,看模型的整体训练性能如何。特别是判断训练是否过拟合的时候非常有效。因此,代码中包含了这部分内容。

       好了,代码终于来了。大家可以搞点大数据库来测试,如果有MPI集群条件的话就更好了。

       下面的代码依赖numpy、scipy(用其稀疏矩阵加速计算)和mpi4py。其中mpi4py需要依赖openmpi和Cpython,可以参考我之前的博客进行安装。

  1. #***************************************************************************
  2. #*
  3. #* Description: label propagation
  4. #* Author: Zou Xiaoyi ([email protected])
  5. #* Date: 2015-10-15
  6. #* HomePage: http://blog.csdn.net/zouxy09
  7. #*
  8. #**************************************************************************
  9. import os, sys, time
  10. import numpy as np
  11. from scipy.sparse import csr_matrix, lil_matrix, eye
  12. import operator
  13. import cPickle as pickle
  14. import mpi4py.MPI as MPI
  15. #
  16. # Global variables for MPI
  17. #
  18. # instance for invoking MPI related functions
  19. comm = MPI.COMM_WORLD
  20. # the node rank in the whole community
  21. comm_rank = comm.Get_rank()
  22. # the size of the whole community, i.e., the total number of working nodes in the MPI cluster
  23. comm_size = comm.Get_size()
  24. # load mnist dataset
  25. def load_MNIST():
  26. import gzip
  27. f = gzip.open("mnist.pkl.gz", "rb")
  28. train, val, test = pickle.load(f)
  29. f.close()
  30. Mat_Label = train[0]
  31. labels = train[1]
  32. Mat_Unlabel = test[0]
  33. groundtruth = test[1]
  34. labels_id = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
  35. return Mat_Label, labels, labels_id, Mat_Unlabel, groundtruth
  36. # return k neighbors index
  37. def navie_knn(dataSet, query, k):
  38. numSamples = dataSet.shape[0]
  39. ## step 1: calculate Euclidean distance
  40. diff = np.tile(query, (numSamples, 1)) - dataSet
  41. squaredDiff = diff ** 2
  42. squaredDist = np.sum(squaredDiff, axis = 1) # sum is performed by row
  43. ## step 2: sort the distance
  44. sortedDistIndices = np.argsort(squaredDist)
  45. if k > len(sortedDistIndices):
  46. k = len(sortedDistIndices)
  47. return sortedDistIndices[0:k]
  48. # build a big graph (normalized weight matrix)
  49. # sparse U x (U + L) matrix
  50. def buildSubGraph(Mat_Label, Mat_Unlabel, knn_num_neighbors):
  51. num_unlabel_samples = Mat_Unlabel.shape[0]
  52. data = []; indices = []; indptr = [0]
  53. Mat_all = np.vstack((Mat_Label, Mat_Unlabel))
  54. values = np.ones(knn_num_neighbors, np.float32) / knn_num_neighbors
  55. for i in xrange(num_unlabel_samples):
  56. k_neighbors = navie_knn(Mat_all, Mat_Unlabel[i, :], knn_num_neighbors)
  57. indptr.append(np.int32(indptr[-1]) + knn_num_neighbors)
  58. indices.extend(k_neighbors)
  59. data.append(values)
  60. return csr_matrix((np.hstack(data), indices, indptr))
  61. # build a big graph (normalized weight matrix)
  62. # sparse U x (U + L) matrix
  63. def buildSubGraph_MPI(Mat_Label, Mat_Unlabel, knn_num_neighbors):
  64. num_unlabel_samples = Mat_Unlabel.shape[0]
  65. local_data = []; local_indices = []; local_indptr = [0]
  66. Mat_all = np.vstack((Mat_Label, Mat_Unlabel))
  67. values = np.ones(knn_num_neighbors, np.float32) / knn_num_neighbors
  68. sample_offset = np.linspace(0, num_unlabel_samples, comm_size + 1).astype('int')
  69. for i in range(sample_offset[comm_rank], sample_offset[comm_rank+1]):
  70. k_neighbors = navie_knn(Mat_all, Mat_Unlabel[i, :], knn_num_neighbors)
  71. local_indptr.append(np.int32(local_indptr[-1]) + knn_num_neighbors)
  72. local_indices.extend(k_neighbors)
  73. local_data.append(values)
  74. data = np.hstack(comm.allgather(local_data))
  75. indices = np.hstack(comm.allgather(local_indices))
  76. indptr_tmp = comm.allgather(local_indptr)
  77. indptr = []
  78. for i in range(len(indptr_tmp)):
  79. if i == 0:
  80. indptr.extend(indptr_tmp[i])
  81. else:
  82. last_indptr = indptr[-1]
  83. del(indptr[-1])
  84. indptr.extend(indptr_tmp[i] + last_indptr)
  85. return csr_matrix((np.hstack(data), indices, indptr), dtype = np.float32)
  86. # label propagation
  87. def run_label_propagation_sparse(knn_num_neighbors = 20, max_iter = 100, tol = 1e-4, test_per_iter = 1):
  88. # load data and graph
  89. print "Processor %d/%d loading graph file..." % (comm_rank, comm_size)
  90. #Mat_Label, labels, Mat_Unlabel, groundtruth = loadFourBandData()
  91. Mat_Label, labels, labels_id, Mat_Unlabel, unlabel_data_id = load_MNIST()
  92. if comm_size > len(labels_id):
  93. raise ValueError("Sorry, the processors must be less than the number of classes")
  94. #affinity_matrix = buildSubGraph(Mat_Label, Mat_Unlabel, knn_num_neighbors)
  95. affinity_matrix = buildSubGraph_MPI(Mat_Label, Mat_Unlabel, knn_num_neighbors)
  96. # get some parameters
  97. num_classes = len(labels_id)
  98. num_label_samples = len(labels)
  99. num_unlabel_samples = Mat_Unlabel.shape[0]
  100. affinity_matrix_UL = affinity_matrix[:, 0:num_label_samples]
  101. affinity_matrix_UU = affinity_matrix[:, num_label_samples:num_label_samples+num_unlabel_samples]
  102. if comm_rank == 0:
  103. print "Have %d labeled images, %d unlabeled images and %d classes" % (num_label_samples, num_unlabel_samples, num_classes)
  104. # divide label_function_U and label_function_L to all processors
  105. class_offset = np.linspace(0, num_classes, comm_size + 1).astype('int')
  106. # initialize local label_function_U
  107. local_start_class = class_offset[comm_rank]
  108. local_num_classes = class_offset[comm_rank+1] - local_start_class
  109. local_label_function_U = eye(num_unlabel_samples, local_num_classes, 0, np.float32, format='csr')
  110. # initialize local label_function_L
  111. local_label_function_L = lil_matrix((num_label_samples, local_num_classes), dtype = np.float32)
  112. for i in xrange(num_label_samples):
  113. class_off = int(labels[i]) - local_start_class
  114. if class_off >= 0 and class_off < local_num_classes:
  115. local_label_function_L[i, class_off] = 1.0
  116. local_label_function_L = local_label_function_L.tocsr()
  117. local_label_info = affinity_matrix_UL.dot(local_label_function_L)
  118. print "Processor %d/%d has to process %d classes..." % (comm_rank, comm_size, local_label_function_L.shape[1])
  119. # start to propagation
  120. iter = 1; changed = 100.0;
  121. evaluation(num_unlabel_samples, local_start_class, local_label_function_U, unlabel_data_id, labels_id)
  122. while True:
  123. pre_label_function = local_label_function_U.copy()
  124. # propagation
  125. local_label_function_U = affinity_matrix_UU.dot(local_label_function_U) + local_label_info
  126. # check converge
  127. local_changed = abs(pre_label_function - local_label_function_U).sum()
  128. changed = comm.reduce(local_changed, root = 0, op = MPI.SUM)
  129. status = 'RUN'
  130. test = False
  131. if comm_rank == 0:
  132. if iter % 1 == 0:
  133. norm_changed = changed / (num_unlabel_samples * num_classes)
  134. print "---> Iteration %d/%d, changed: %f" % (iter, max_iter, norm_changed)
  135. if iter >= max_iter or changed < tol:
  136. status = 'STOP'
  137. print "************** Iteration over! ****************"
  138. if iter % test_per_iter == 0:
  139. test = True
  140. iter += 1
  141. test = comm.bcast(test if comm_rank == 0 else None, root = 0)
  142. status = comm.bcast(status if comm_rank == 0 else None, root = 0)
  143. if status == 'STOP':
  144. break
  145. if test == True:
  146. evaluation(num_unlabel_samples, local_start_class, local_label_function_U, unlabel_data_id, labels_id)
  147. evaluation(num_unlabel_samples, local_start_class, local_label_function_U, unlabel_data_id, labels_id)
  148. def evaluation(num_unlabel_samples, local_start_class, local_label_function_U, unlabel_data_id, labels_id):
  149. # get local label with max score
  150. if comm_rank == 0:
  151. print "Start to combine local result..."
  152. local_max_score = np.zeros((num_unlabel_samples, 1), np.float32)
  153. local_max_label = np.zeros((num_unlabel_samples, 1), np.int32)
  154. for i in xrange(num_unlabel_samples):
  155. local_max_label[i, 0] = np.argmax(local_label_function_U.getrow(i).todense())
  156. local_max_score[i, 0] = local_label_function_U[i, local_max_label[i, 0]]
  157. local_max_label[i, 0] += local_start_class
  158. # gather the results from all the processors
  159. if comm_rank == 0:
  160. print "Start to gather results from all processors"
  161. all_max_label = np.hstack(comm.allgather(local_max_label))
  162. all_max_score = np.hstack(comm.allgather(local_max_score))
  163. # get terminate label of unlabeled data
  164. if comm_rank == 0:
  165. print "Start to analysis the results..."
  166. right_predict_count = 0
  167. for i in xrange(num_unlabel_samples):
  168. if i % 1000 == 0:
  169. print "***", all_max_score[i]
  170. max_idx = np.argmax(all_max_score[i])
  171. max_label = all_max_label[i, max_idx]
  172. if int(unlabel_data_id[i]) == int(labels_id[max_label]):
  173. right_predict_count += 1
  174. accuracy = float(right_predict_count) * 100.0 / num_unlabel_samples
  175. print "Have %d samples, accuracy: %.3f%%!" % (num_unlabel_samples, accuracy)
  176. if __name__ == '__main__':
  177. run_label_propagation_sparse(knn_num_neighbors = 20, max_iter = 30)

五、参考资料

[1]Semi-SupervisedLearning with Graphs.pdf