python - 将标签映射到 tf.unique_with_counts() 之后的计数

Question

我有一个 BxN 输入特征张量 feature_labels，我调用

unique_labels, label_idcs, label_counts = tf.unique_with_counts(feature_labels)

获取标签计数。

如何生成 BxN 矩阵，用于存储标签计数而不是标签？

一批示例:

Input: 
feature_labels: 0 1 1 1 0 2 3 1 3

> unique_labels, _, label_counts = tf.unique_with_counts(feature_labels)

=> unique_labels: 0 1 2 3
=> label_counts: 2 4 1 2

Output:
2 4 4 4 2 1 2 4 2

Answer 1

如果我理解正确，我认为你可以这样做:

import tensorflow as tf

with tf.Graph().as_default(), tf.Session() as sess:
    feature_labels = tf.constant([0, 1, 1, 1, 0, 2, 3, 1, 3], tf.int32)
    _, unique_idx, label_counts = tf.unique_with_counts(feature_labels)
    result = tf.gather(label_counts, unique_idx)
    print(sess.run(result))
    # [2 4 4 4 2 1 2 4 2]