Question

我偶然发现了这件事，但无法理解发生的事情。我正在对图像进行K-means聚类分配，并尝试对代码进行矢量化以使其尽可能快地运行。我想出了以下代码：

image_values =np.array( [[[ 0.36302522  0.51708686  0.20952381]
                      [ 0.46330538  0.69915968  0.2140056 ]
                      [ 0.7904762   0.93837535  0.27002802]
                      [ 0.78375351  0.89187676  0.24201682]
                      [ 0.57871151  0.79775912  0.24593839]
                      [ 0.2896359   0.39103645  0.64481789]
                      [ 0.23809525  0.30924368  0.64257705]]

                     [[ 0.36302522  0.51708686  0.20952381]
                      [ 0.46330538  0.69915968  0.2140056 ]
                      [ 0.7904762   0.93837535  0.27002802]
                      [ 0.78375351  0.89187676  0.24201682]
                      [ 0.57871151  0.79775912  0.24593839]
                      [ 0.2896359   0.39103645  0.64481789]
                      [ 0.23809525  0.30924368  0.64257705]]

                     [[ 0.36302522  0.51708686  0.20952381]
                      [ 0.46330538  0.69915968  0.2140056 ]
                      [ 0.7904762   0.93837535  0.27002802]
                      [ 0.78375351  0.89187676  0.24201682]
                      [ 0.57871151  0.79775912  0.24593839]
                      [ 0.2896359   0.39103645  0.64481789]
                      [ 0.23809525  0.30924368  0.64257705]]])
means = np.array([[0.909,0.839,0.6509],[0.813,0.808,0.694],[0.331,0.407,0.597]]) #random centroids
err = 1
while err > .01:
    J = [np.sum((image_values-avg)**2, axis = 2) for avg in means]
    K = np.argmin(J, axis = 0)
    old_means = means
    means = np.array([np.mean(image_values[K==i], axis ==True) for i in range(len(means))])
    print means
    err = abs(sum(old_means)-sum(means))
    print err

在每个新的平均值计算中，我使用我的K数组来选择每个平均值计算中应包含哪些像素值，但我无法让轴达成一致。我实际上做了一个错字，而不是轴= 3，我键入轴== 3，它工作！我尝试了一堆不同的数字，发现数字是什么并不重要，结果是一样的。我尝试了一堆数字和布尔运算器，他们没有工作。我已经阅读了文档，但我无法理解。

numpy在其中一个数组函数的axis参数中获取逻辑if时会做什么？

谢谢！

Answer 1

我不完全确定我完全明白你要做什么。这就是我的假设;您有一个具有RGB值的单个图像，并且您希望对此图像中的像素进行聚类。因此，每个质心将分别为每个颜色通道定义一个值。我假设means矩阵中的每一行都是一个质心，列是RGB值。

在你的方法中，我认为你减去质心的方式可能有误。您需要为每个质心创建一个距离矩阵（此时您没有正确地减去每个颜色通道）。

这是一个主张。请注意，对于给定的示例数据，您将遇到NaN错误，因为并非所有质心都具有最接近它们的像素。您可能还需要根据需要调整停止标准。

err = 1
while err > 0.1:
    # There are three centroids. We would like to compute the
    # distance for each pixel to each centroid. Here, the image
    # is thus replicated three times.
    dist = np.tile(image_values, (3,1,1,1))

    # The 2D matrix needs to be reshaped to fit the dimensions of
    # the dist matrix. With the new shape, the matrix can directly
    # be subtracted.
    means2 = means.reshape(3,3,1,1)

    # Subtract each respective RGB value of the centroid for 
    # each "replica" of the image
    J = np.power(dist - means2, 2)

    # Sum the r,g,b channels together to get the total distance for a pixel
    J = J.sum(axis=1)

    # Check for which cluster the pixel is closest
    K = np.argmin(J, axis=0)

    # I couldn't think of a better way than this loop
    newMeans = np.zeros((3,3))
    for i in range(means.shape[0]): # do each centroid
        # In axis 1 there are pixels which we would like to
        # average for each color channel (axis 0 are the RGB channels)
        newMeans[i,:] = image_values[:,K==i].mean(axis=1)

    err = np.power(means - newMeans, 2).sum()
    means = newMeans

numpy函数如何处理axis参数的逻辑if运算符？

1 个答案: