关于 tensorflow 1.x sparse_tensor_dense_matmul 的问题

Question

tf.sparse_tensor_dense_matmul 是对稀疏张量和稠密张量做矩阵乘法的函数，但为什么我运算的结果与 numpy 运算结果不同？

import numpy as np
import tensorflow as tf
import scipy.sparse as sp

arr = np.array([[1.,2.,5.,2.],[3.,4.,1.,2.],[3.,5.,2.,6.],[4.,13.,2.,10.]]) # 4*4
arr = sp.coo_matrix(arr)

b = np.array([[1.],[2.],[3.],[4.]])
b = tf.convert_to_tensor(b, dtype=tf.float32)


if not sp.isspmatrix_coo(arr):
    arr = arr.tocoo()

arr = arr.astype(np.float32)
indices = np.vstack((arr.col, arr.row)).transpose()


a_sp = tf.SparseTensor(indices=indices, values=arr.data, dense_shape=arr.shape)

with tf.Session() as sess:
    m=tf.sparse_tensor_dense_matmul(a_sp, b)
    print(sess.run(m))

结果：

[[32.]
 [77.]
 [21.]
 [64.]]

numpy 下：

x = np.array([[1,2,5,2],[3,4,1,2],[3,5,2,6],[4,13,2,10]],dtype=np.float)
y = np.array([[1],[2],[3],[4]], dtype=np.float)
x @ y

结果：

array([[28.],
       [22.],
       [43.],
       [76.]]

Answer 1

同学，你的indices搞反了

indices = np.vstack(( arr.row, arr.col)).transpose()