如何在2D张量的第二维上使用Tensorflows scatter_nd? [英] How can I use Tensorflows scatter_nd on the second dimension of a 2D tensor?

问题描述

TL; DR:如何将每个实例2个标签的2D二进制张量拆分为每个实例仅1个标签的2个张量，如下图所示:

TL;DR: How can I split 2D binary tensor of 2 labels per instance, into 2 tensors with only 1 label per instance, like in this pic:

作为自定义损失函数的一部分，我试图将每个实例2个标签的多标签y张量拆分为每个实例1个标签的2个y张量. 当我在1D y张量上执行此代码时，此代码非常有用:

As part of a custom loss function, I'm trying to split a multi-label y tensor, with 2 labels per instance, to 2 y tensors with 1 label per instance. When I'm doing it on 1D y tensor, this code works great:

y_true = tf.constant([1., 0., 0., 0., 1., 0., 0., 0., 0.])
label_cls = tf.where(tf.equal(y_true, 1.))
idx1, idx2 = tf.split(label_cls,2)
raplace = tf.constant([1.])
y_true_1 = tf.scatter_nd(tf.cast(idx1, dtype=tf.int32), raplace, [tf.size(y_true)]) 
y_true_2 = tf.scatter_nd(tf.cast(idx2, dtype=tf.int32), raplace, [tf.size(y_true)])  

with tf.Session() as sess:
    print(sess.run([y_true_1,y_true_2]))

然后我得到:

[array([1., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32), array([0., 0., 0., 0., 1., 0., 0., 0., 0.], dtype=float32)]

但是当我在训练中使用批处理时，会出现以下错误:

But when I use batches in training, I get this error:

Invalid argument: Outer dimensions of indices and update must match.

由于我的"y张量"是2D而不是1D，因此在这种情况下-idx1, idx2(索引)不正确，replace的形状(更新)也不正确. 据我了解，tf.scatter_nd只能更新变量的第一维，那么如何解决呢?以及如何获取所需的索引?

Since my "y tensors" are 2D and not 1D, and in this case- idx1, idx2 (the indices) are not right, nor do the shape of replace (the updates). For what I understand, tf.scatter_nd can only update the first dimension of the variable, so how can I get around it? and how can I get the needed indices for that?

推荐答案

我觉得您走的是曲折的道路.这是我的解决方案.感觉比尝试的方法更简单(tf 1.14尝试).

I feel like you're taking a convoluted path. Here's my solution. Feel it's more straightforward than the one you're trying to go with (Attempted with tf 1.14).

import tensorflow as tf

y_true = tf.constant([[1, 0, 1, 0],[0, 1, 1, 0]])
_, label_inds = tf.math.top_k(y_true, k=2)
idx1, idx2 = tf.split(label_inds,2, axis=1)

y_true_1 = tf.one_hot(idx1, depth=4)
y_true_2 = tf.one_hot(idx2, depth=4)

with tf.Session() as sess:

    print(sess.run([y_true_1, y_true_2]))

因此，您的想法是获取每行的前2个标签的索引.然后使用tf.split将其分为2列.然后使用one_hot将这些索引转换回一个热向量.

So the idea is that you get the indices of top 2 labels for each row. Then split that into 2 columns using tf.split. And then use one_hot to convert those indices back to onehot vectors.

这篇关于如何在2D张量的第二维上使用Tensorflows scatter_nd?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！