我将如何使用 TensorFlow 实现 k-means?
Posted
技术标签:
【中文标题】我将如何使用 TensorFlow 实现 k-means?【英文标题】:How would I implement k-means with TensorFlow? 【发布时间】:2016-02-10 20:06:27 【问题描述】:使用内置梯度下降优化器的介绍教程很有意义。然而,k-means 不仅仅是我可以插入梯度下降的东西。似乎我必须编写自己的优化器,但考虑到 TensorFlow 原语,我不太确定如何做到这一点。
我应该采取什么方法?
【问题讨论】:
【参考方案1】:现在您可以直接使用(或从中获取灵感)KMeansClustering Estimator。你可以看看its implementation on GitHub。
【讨论】:
【参考方案2】:到目前为止,我看到的大多数答案都只关注 2d 版本(当您需要对 2 维中的点进行聚类时)。这是我在任意维度上的聚类实现。
k-means algorithm in n dims 的基本思想:
随机生成k个起点 执行此操作,直到您超出耐心或集群分配没有改变: 将每个点分配给最近的起点 通过取其集群中的平均值来重新计算每个起点的位置为了能够以某种方式验证结果,我将尝试对 MNIST 图像进行聚类。
import numpy as np
import tensorflow as tf
from random import randint
from collections import Counter
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/")
X, y, k = mnist.test.images, mnist.test.labels, 10
所以这里X是我要聚类(10000, 784)
的数据,y是实数,k是聚类数(与位数相同。现在实际算法:
# select random points as a starting position. You can do better by randomly selecting k points.
start_pos = tf.Variable(X[np.random.randint(X.shape[0], size=k),:], dtype=tf.float32)
centroids = tf.Variable(start_pos.initialized_value(), 'S', dtype=tf.float32)
# populate points
points = tf.Variable(X, 'X', dtype=tf.float32)
ones_like = tf.ones((points.get_shape()[0], 1))
prev_assignments = tf.Variable(tf.zeros((points.get_shape()[0], ), dtype=tf.int64))
# find the distance between all points: http://***.com/a/43839605/1090562
p1 = tf.matmul(
tf.expand_dims(tf.reduce_sum(tf.square(points), 1), 1),
tf.ones(shape=(1, k))
)
p2 = tf.transpose(tf.matmul(
tf.reshape(tf.reduce_sum(tf.square(centroids), 1), shape=[-1, 1]),
ones_like,
transpose_b=True
))
distance = tf.sqrt(tf.add(p1, p2) - 2 * tf.matmul(points, centroids, transpose_b=True))
# assign each point to a closest centroid
point_to_centroid_assignment = tf.argmin(distance, axis=1)
# recalculate the centers
total = tf.unsorted_segment_sum(points, point_to_centroid_assignment, k)
count = tf.unsorted_segment_sum(ones_like, point_to_centroid_assignment, k)
means = total / count
# continue if there is any difference between the current and previous assignment
is_continue = tf.reduce_any(tf.not_equal(point_to_centroid_assignment, prev_assignments))
with tf.control_dependencies([is_continue]):
loop = tf.group(centroids.assign(means), prev_assignments.assign(point_to_centroid_assignment))
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# do many iterations. Hopefully you will stop because of has_changed is False
has_changed, cnt = True, 0
while has_changed and cnt < 300:
cnt += 1
has_changed, _ = sess.run([is_continue, loop])
# see how the data is assigned
res = sess.run(point_to_centroid_assignment)
现在是时候检查我们的集群有多好。为此,我们将集群中出现的所有实数组合在一起。之后,我们将看到该集群中最受欢迎的选择。在完美聚类的情况下,我们将在每组中只有一个值。在随机簇的情况下,每个值将在组中大致相等。
nums_in_clusters = [[] for i in xrange(10)]
for cluster, real_num in zip(list(res), list(y)):
nums_in_clusters[cluster].append(real_num)
for i in xrange(10):
print Counter(nums_in_clusters[i]).most_common(3)
这给了我这样的东西:
[(0, 738), (6, 18), (2, 11)]
[(1, 641), (3, 53), (2, 51)]
[(1, 488), (2, 115), (7, 56)]
[(4, 550), (9, 533), (7, 280)]
[(7, 634), (9, 400), (4, 302)]
[(6, 649), (4, 27), (0, 14)]
[(5, 269), (6, 244), (0, 161)]
[(8, 646), (5, 164), (3, 125)]
[(2, 698), (3, 34), (7, 14)]
[(3, 712), (5, 290), (8, 110)]
这很好,因为大多数计数都在第一组中。您会看到聚类混淆了 7 和 9、4 和 5。但 0 的聚类效果非常好。
一些改进方法:
多次运行该算法并选择最佳算法(基于到集群的距离) 处理没有分配给集群的情况。在我的例子中,你会在means
变量中得到 Nan,因为 count
是 0。
随机点初始化。
【讨论】:
【参考方案3】:(注意:您现在可以获取a more polished version of this code as a gist on github。)
您绝对可以这样做,但您需要定义自己的优化标准(对于 k-means,它通常是最大迭代次数以及分配稳定的时间)。这是一个示例,说明您可以如何做到这一点(可能有更优化的方法来实现它,当然还有更好的方法来选择初始点)。如果你真的很努力地远离在 python 中迭代地做事情,基本上就像你会在 numpy 中做这件事:
import tensorflow as tf
import numpy as np
import time
N=10000
K=4
MAX_ITERS = 1000
start = time.time()
points = tf.Variable(tf.random_uniform([N,2]))
cluster_assignments = tf.Variable(tf.zeros([N], dtype=tf.int64))
# Silly initialization: Use the first two points as the starting
# centroids. In the real world, do this better.
centroids = tf.Variable(tf.slice(points.initialized_value(), [0,0], [K,2]))
# Replicate to N copies of each centroid and K copies of each
# point, then subtract and compute the sum of squared distances.
rep_centroids = tf.reshape(tf.tile(centroids, [N, 1]), [N, K, 2])
rep_points = tf.reshape(tf.tile(points, [1, K]), [N, K, 2])
sum_squares = tf.reduce_sum(tf.square(rep_points - rep_centroids),
reduction_indices=2)
# Use argmin to select the lowest-distance point
best_centroids = tf.argmin(sum_squares, 1)
did_assignments_change = tf.reduce_any(tf.not_equal(best_centroids,
cluster_assignments))
def bucket_mean(data, bucket_ids, num_buckets):
total = tf.unsorted_segment_sum(data, bucket_ids, num_buckets)
count = tf.unsorted_segment_sum(tf.ones_like(data), bucket_ids, num_buckets)
return total / count
means = bucket_mean(points, best_centroids, K)
# Do not write to the assigned clusters variable until after
# computing whether the assignments have changed - hence with_dependencies
with tf.control_dependencies([did_assignments_change]):
do_updates = tf.group(
centroids.assign(means),
cluster_assignments.assign(best_centroids))
sess = tf.Session()
sess.run(tf.initialize_all_variables())
changed = True
iters = 0
while changed and iters < MAX_ITERS:
iters += 1
[changed, _] = sess.run([did_assignments_change, do_updates])
[centers, assignments] = sess.run([centroids, cluster_assignments])
end = time.time()
print ("Found in %.2f seconds" % (end-start)), iters, "iterations"
print "Centroids:"
print centers
print "Cluster assignments:", assignments
(请注意,真正的实现需要更加小心初始集群选择,避免所有点都进入一个集群等问题。这只是一个快速演示。我已经更新了我之前的答案它更清晰和“值得举例”。)
【讨论】:
我应该解释得更好一点。它需要 N 个点并制作它们的 K 个副本。它需要 K 个当前质心并制作 N 个副本。然后它减去这两个大张量以获得从每个点到每个质心的 N*K 距离。它计算距离的平方和,并使用“argmin”为每个点找到最佳距离。然后它使用 dynamic_partition 根据它们的聚类分配将点分组到 K 个不同的张量中,找到每个聚类中的均值,并根据它设置质心。以上是关于我将如何使用 TensorFlow 实现 k-means?的主要内容,如果未能解决你的问题,请参考以下文章
TensorFlow:如何使用具有可变输入长度的 CudnnLSTM(如 dynamic_rnn)?