TensorFlow - 类似 numpy 的张量索引 [英] TensorFlow - numpy-like tensor indexing

问题描述

在 numpy 中，我们可以这样做:

In numpy, we can do this:

x = np.random.random((10,10))
a = np.random.randint(0,10,5)
b = np.random.randint(0,10,5)
x[a,b] # gives 5 entries from x, indexed according to the corresponding entries in a and b

当我在 TensorFlow 中尝试类似的东西时:

When I try something equivalent in TensorFlow:

xt = tf.constant(x)
at = tf.constant(a)
bt = tf.constant(b)
xt[at,bt]

最后一行给出了Bad slice index tensor"异常.TensorFlow 似乎不支持像 numpy 或 Theano 这样的索引.

The last line gives a "Bad slice index tensor" exception. It seems TensorFlow doesn't support indexing like numpy or Theano.

有谁知道是否有一种 TensorFlow 方法可以做到这一点(通过任意值索引张量).我已经看过 tf.nn.embedding 部分，但我不确定它们是否可以用于此目的，即使可以，对于如此简单的事情来说，这也是一个巨大的解决方法.

Does anybody know if there is a TensorFlow way of doing this (indexing a tensor by arbitrary values). I've seen the tf.nn.embedding part, but I'm not sure they can be used for this and even if they can, it's a huge workaround for something this straightforward.

(现在，我将来自 x 的数据作为输入并在 numpy 中进行索引，但我希望将 x 放在 TensorFlow 中以获得更高的效率)

(Right now, I'm feeding the data from x as an input and doing the indexing in numpy but I hoped to put x inside TensorFlow to get higher efficiency)

推荐答案

您现在实际上可以使用 tf.gather_nd.假设您有一个矩阵 m，如下所示:

You can actually do that now with tf.gather_nd. Let's say you have a matrix m like the following:

| 1 2 3 4 |
| 5 6 7 8 |

并且您想要构建一个大小为 r 的矩阵，比方说，3x2，由 m 的元素构建，如下所示:

And you want to build a matrix r of size, let's say, 3x2, built from elements of m, like this:

| 3 6 |
| 2 7 |
| 5 3 |
| 1 1 |

r的每个元素对应m的一行和一列，可以有矩阵rows和cols 带有这些索引(从零开始，因为我们正在编程，而不是在做数学！):

Each element of r corresponds to a row and column of m, and you can have matrices rows and cols with these indices (zero-based, since we are programming, not doing math!):

       | 0 1 |         | 2 1 |
rows = | 0 1 |  cols = | 1 2 |
       | 1 0 |         | 0 2 |
       | 0 0 |         | 0 0 |

你可以像这样堆叠成一个 3 维张量:

Which you can stack into a 3-dimensional tensor like this:

| | 0 2 | | 1 1 | |
| | 0 1 | | 1 2 | |
| | 1 0 | | 2 0 | |
| | 0 0 | | 0 0 | |

这样就可以通过rows和cols从m到r，如下:

This way, you can get from m to r through rows and cols as follows:

import numpy as np
import tensorflow as tf

m = np.array([[1, 2, 3, 4], [5, 6, 7, 8]])
rows = np.array([[0, 1], [0, 1], [1, 0], [0, 0]])
cols = np.array([[2, 1], [1, 2], [0, 2], [0, 0]])

x = tf.placeholder('float32', (None, None))
idx1 = tf.placeholder('int32', (None, None))
idx2 = tf.placeholder('int32', (None, None))
result = tf.gather_nd(x, tf.stack((idx1, idx2), -1))

with tf.Session() as sess:
    r = sess.run(result, feed_dict={
        x: m,
        idx1: rows,
        idx2: cols,
    })
print(r)

输出:

[[ 3.  6.]
 [ 2.  7.]
 [ 5.  3.]
 [ 1.  1.]]

这篇关于TensorFlow - 类似 numpy 的张量索引的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！