如何使用 tf.scatter_add 在 tensorflow 中增加矩阵元素? [英] how to increment matrix element in tensorflow using tf.scatter_add?

问题描述

tf.scatter_add 适用于 1d(形状 1) 张量:

tf.scatter_add works nicely for 1d (shape 1) tensors:

> S = tf.Variable(tf.constant([1,2,3,4]))
> sess.run(tf.initialize_all_variables())
> sess.run(tf.scatter_add(S, [0], [10]))

array([11,  2,  3,  4], dtype=int32)

> sess.run(tf.scatter_add(S, [0, 1], [10, 100]))

array([ 21, 102,   3,   4], dtype=int32)

但是我怎样才能增加，比如说

But how can I increment, say [0,0] element of

M = tf.Variable(tf.constant([[1,2], [3,4]]))

使其成为 [[2, 2], [3, 4]]使用 tf.scatter_add?

to make it [[2, 2], [3, 4]] using tf.scatter_add?

官方文档很友好'一个神秘的.我尝试了不同的 arg 值，比如

the official documentation is kind'a cryptic. And I tried different arg values, say

> sess.run(tf.scatter_add(M, [[0, 0]], [1]))
*** ValueError: Shapes (1,) and (1, 2, 2) are not compatible

并没有成功.

顺便说一句，就我而言，M 非常大并且可以动态调整大小.因此，将零一等于一元素矩阵添加到 M 不是这种情况.

Btw, in my case, M is quite large and is resized dynamically. So adding zero-but-one equal to 1 element matrix to M is not the case.

推荐答案

tf.scatter_add 更新张量的切片，但不能更新单个系数.例如，它可以一次更新矩阵的整行.

tf.scatter_add updates slices of a tensor and not capable of updating individual coefficients. For instance, it can update entire rows of a matrix at once.

此外，tf.scatter_add 的 updates 参数的形状取决于其 indices 参数的形状.当 ref 参数是一个形状为 (M, N) 的矩阵时，则

Also, the shape of the updates argument to tf.scatter_add depends on the shape of its indices argument. When the ref argument is a matrix with shape (M, N), then

• 如果indices 是一个标量i，那么updates 应该是一个形状为(N) 的向量.
• 如果indices是一个形状为(k)的向量[i1, i2, .. ik]，则updates 应该具有 (k, N) 的形状.

• If indices is a scalar i, then updates should be a vector with shape (N).
• If indices is a vector [i1, i2, .. ik] with shape (k), then updates should have the shape (k, N).

在你的情况下，你可以简单地将 [1, 0] 添加到 M 的第一行，如下所示，以获得你想要的效果:

In your case, you can simply add [1, 0] to the first row of M as follows to get the effect you want:

sess.run(tf.scatter_add(M, 0, [1, 0]))
array([[2, 2],
   [3, 4]], dtype=int32)

这篇关于如何使用 tf.scatter_add 在 tensorflow 中增加矩阵元素?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！