Tensorflow:将输出张量转换为 one-hot [英] Tensorflow: Convert output tensor to one-hot

问题描述

我想做什么:(来自 cs231n 冬季课程)

What I want to do: (from cs231n Winter course)

我将使用 tensorflow 来实现这个.

I'm gonna implement this using tensorflow.

但问题是我不知道如何将分数转换为 one-hot(上图中的红色线条)

But problem is I have no idea how to convert scores to one-hot (red colored line in above image)

假设我有一个 model 类，它具有所有张量操作作为对象变量.

Let say I have a model class which has all tensor operations as object variables.

model.outputs 是获取 scores 的张量操作(前馈)，我需要将此 outputs 张量转换为 one-hot张量以不同的方式，以便我可以执行梯度操作.

model.outputs is an tensor operation(feedforward) to get the scores and I need to convert this outputs tensor to one-hot tensor IN A DIFFERENCIABLE WAY so that I can perform the gradient operation.

我该如何实施?

推荐答案

假设您的 scores 或 model.outputs 节点的形状为 [batch, #类].在下面的例子中，我们使用 batch_size = 2，我们有 4 个 classes.

Suppose your scores or model.outputs node has a shape of [batch, #class]. In the example below, we use batch_size = 2, and we have 4 classes.

tf.reset_default_graph()
batch_size=2
num_classes=4
score = tf.constant([[0.5, 0.6, 0.2, 0.01], 
                     [0.8, 0.75, 1.0, 1.0]])

max_per_instance = tf.expand_dims(tf.reduce_max(score, axis=1), 0)
tiled = tf.tile(max_per_instance, [num_classes, 1])
n_hot = tf.cast(tf.equal(score, tf.transpose(tiled)), tf.int32)
with tf.Session() as sess:
  print sess.run(n_hot)

>> [[0 1 0 0]
    [0 0 1 1]]

这篇关于Tensorflow:将输出张量转换为 one-hot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！