Tensorflow 高级索引：根据两个索引张量将较小的张量分配到较大的张量中

Question

我想创建一个形状为 (3,4,2,2) 的 zeros 张量，并在两个 (3,1) 张量给定的位置插入一个 (3,4) 张量。

示例代码：对数组的等效 numpy 操作如下：

# Existing arrays of required shapes
bbox = np.arange(3*4).reshape(3,4)
x = np.array([0,0,1])
y = np.array([1,1,1])

# Create zeros array and assign into it
output = np.zeros((3,4,2,2))
output[np.arange(3),:,x,y] = bbox

如何使用 Tensorflow 做类似的事情？

注意：我实际上想使用大小为 (32,125,32,32) 的张量。以上是简单的复制代码

Answer 1

以下是使用tf.scatter_nd 的方法：

import tensorflow as tf
import numpy as np

bbox = np.arange(3 * 4).reshape(3, 4)
x = np.array([0, 0, 1])
y = np.array([1, 1, 1])
x_size = 2
y_size = 2

# TensorFlow calculation
with tf.Graph().as_default(), tf.Session() as sess:
    bbox_t = tf.convert_to_tensor(bbox)
    x_t = tf.convert_to_tensor(x)
    y_t = tf.convert_to_tensor(y)
    shape = tf.shape(bbox_t)
    rows, cols = shape[0], shape[1]
    ii, jj = tf.meshgrid(tf.range(rows), tf.range(cols), indexing='ij')
    xx = tf.tile(tf.expand_dims(x_t, 1), (1, cols))
    yy = tf.tile(tf.expand_dims(y_t, 1), (1, cols))
    idx = tf.stack([ii, jj, xx, yy], axis=-1)
    output = tf.scatter_nd(idx, bbox_t, [rows, cols, x_size, y_size])
    output_tf = sess.run(output)

# Test with NumPy calculation
output_np = np.zeros((3, 4, 2, 2))
output_np[np.arange(3), :, x, y] = bbox
print(np.all(output_tf == output_np))
# True