kMeans Clustering using TF2.0, ValueError:

Question

我正在尝试使用 TensorFlow 2.0 实现简单的 k-means 聚类。预计@tf.function使用 autograph 来转换带有 for 循环的函数。请让我知道导致 ValueError 的原因。tf_kmeans.pyimport tensorflow as tfimport numpy as npfrom typeguard import typecheckedfrom typing import [e-mail] train_kmeans(X: Union[tf.Tensor, np.ndarray],     k: Union[int, tf.Tensor],     n_iter: Union[int, tf.Tensor] = 10) -> (tf.Tensor, tf.Tensor):    X = tf.convert_to_tensor(X)    X = tf.cast(X, tf.float32)    assert len(tf.shape(X)) == 2, "Training data X must be represented as 2D array only"    m = tf.shape(X)[0]    k = tf.convert_to_tensor(k, dtype=tf.int64)    random_select = tf.random.shuffle(X)    init_centroids = random_select[:k]    centroids = tf.Variable(init_centroids)    clusters = tf.zeros([m, ], dtype=tf.int64)    clusters = tf.Variable(clusters)    for _ in tf.range(n_iter):        squared_diffs = tf.square(X[None, :, :] - centroids[:, None, :])        euclidean_dists = tf.reduce_sum(squared_diffs, axis=-1) ** 0.5        clusters.assign(tf.argmin(euclidean_dists, axis=0))        selector = tf.range(k)[:, None] == clusters[None, :]        for c in tf.range(k):            select = selector[c]            points = X[select]            mean_points = tf.reduce_mean(points, axis=0)            centroids[c].assign(mean_points)    centroids = tf.convert_to_tensor(centroids)    return centroids, clusters以下代码用于调用该函数：tf_means_test.pyimport tensorflow as tfimport numpy as npX = np.array([[ 2., 10.],    [ 2.,  5.],    [ 8.,  4.],    [ 5.,  8.],    [ 7.,  5.],    [ 6.,  4.],    [ 1.,  2.],    [ 4.,  9.]])k = 3import tf_kmeanscentroids, clusters = tf_kmeans.train_kmeans(X, k)print(centroids)print(clusters)如果tf.function删除了装饰器，则代码可以正常工作，因为在这种情况下不会执行签名。提前致谢。

Answer 1

我假设您想要一个tf.Variable仅使用的实例assign。但是，在使用 时tf.function，您应该始终从外部提供变量，并在内部使用内置的 TensorFlow 数据结构。

例如，您的代码更改最少，没有tf.Variable对象将是：

import tensorflow as tf

import numpy as np

from typeguard import typechecked

from typing import Union

@tf.function

def train_kmeans(X: Union[tf.Tensor, np.ndarray], 

    k: Union[int, tf.Tensor], 

    n_iter: Union[int, tf.Tensor] = 10) -> (tf.Tensor, tf.Tensor):

    X = tf.convert_to_tensor(X)

    X = tf.cast(X, tf.float32)

    # Required as an int later

    num_centers = k

    assert len(tf.shape(X)) == 2, "Training data X must be represented as 2D array only"

    m = tf.shape(X)[0]

    k = tf.convert_to_tensor(k, dtype=tf.int64)

    random_select = tf.random.shuffle(X)

    init_centroids = random_select[:k]

    centroids = init_centroids

    clusters = tf.zeros([m, ], dtype=tf.int64)

    for _ in tf.range(n_iter):

        squared_diffs = tf.square(X[None, :, :] - centroids[:, None, :])

        euclidean_dists = tf.reduce_sum(squared_diffs, axis=-1) ** 0.5

        clusters = tf.argmin(euclidean_dists, axis=0)

        selector = tf.range(k)[:, None] == clusters[None, :]

        # TF data structure

        new_centroids = tf.TensorArray(tf.float32, num_centers, element_shape=[1, 2])

        for c in range(k):

            select = selector[c]

            points = X[select]

            centroid = tf.reduce_mean(points, axis=0)

            centroid = tf.reshape(centroid, [1, 2])

            new_centroids.write(tf.cast(c, tf.int32), centroid)

        centroids = new_centroids.concat()

        centroids = tf.reshape(centroids, [num_centers, 2])

    return centroids, clusters