TensorFlow 中的索引操作

Question

我在做一些数据的批量标注时，有一个变量用来记录所有的计算结果：

p_all = tf.Variable(tf.zeros([batch_num, batch_size]), name = "probability");

在计算中，我有一个循环来处理每个批次：

for i in range(batch_num):
    feed = {x: testDS.test.next_batch(batch_size)}
    sess.run(p_each_batch, feed_dict=feed)

如何将 p_each_bach 的值复制到 p_all 中？

为了更清楚，我想要这样的东西：

... ...
p_all[batch_index,:] = p_each_batch
for i in range(batch_num):
    feed = {x: testDS.test.next_batch(batch_size), batch_index: i}
    sess.run(p_all, feed_dict=feed)

我怎样才能使这些代码真正起作用？

Answer 1

因为 p_all 是一个 tf.Variable, you can use the tf.scatter_update() 操作来更新每个批次中的单个行：

# Equivalent to `p_all[batch_index, :] = p_each_batch`
update_op = tf.scatter_update(p_all,
                              tf.expand_dims(batch_index, 0),
                              tf.expand_dims(p_each_batch, 0)) 

for i in range(batch_num):
    feed = {x: testDS.test.next_batch(batch_size), batch_index: i}
    sess.run(update_op, feed_dict=feed)

TensorFlow 中的索引操作

Index operation in TensorFlow

indexing

tensorflow