Tensorflow - 张量的块更新

Posted 2023-02-23

【中文标题】Tensorflow - 张量的块更新

【英文标题】：Tensorflow - Block Update of a tensor

【发布时间】：2018-05-13 08:42:47

【问题描述】：

x 是一个形状为 [32,32,3] 的张量

Y 是一个形状为 [1,320,320,3] 的张量

我们可以将 Y, Y[0,:32,:32,:] = x 更新为张量流操作吗？

【参考方案1】：

是的，你可以。你测试过吗？

切片也是张量。使用：

Y[0,:32,:32,:].assign(x)

1D 示例，以便我们清楚地看到正在发生的事情：

import numpy as np
import tensorflow as tf


x = tf.Variable(np.zeros((2,)))
Y = tf.Variable(np.ones((10,)))

block_update = Y[4:6].assign(x)

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    print('old', Y.eval(sess))
    sess.run(block_update)
    print('new', Y.eval(sess))

输出：

old [ 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.]
new [ 1.  1.  1.  1.  0.  0.  1.  1.  1.  1.]

【参考方案2】：

gen_array_ops.inplace_update 可能是您想要的。

from tensorflow.python.ops.gen_array_ops import inplace_update

这里是官方文档

def inplace_update(x, i, v, name=None):
  r"""Updates specified rows 'i' with values 'v'.

  Computes `x[i, :] = v; return x`.

  Originally this function is mutative however for compilation we make this
  operation create / operate on a copy of `x`.

  Args:
    x: A `Tensor`. A tensor of type `T`.
    i: A `Tensor` of type `int32`.
      A vector. Indices into the left-most dimension of `x`.
    v: A `Tensor`. Must have the same type as `x`.
      A `Tensor` of type T. Same dimension sizes as x except the first dimension, which must be the same as i's size.
    name: A name for the operation (optional).

  Returns:
    A `Tensor`. Has the same type as `x`.
  """