为何tensorly.tenalg模块中不存在tensor_dot属性？

问题场景 #

编写了计算三个矩阵Kronecker积的函数matrix_outer，目标是通过Tensorly库的tensor_dot实现张量外推算法，但运行时抛出属性不存在的异常，而官方文档显示该函数存在，需排查原因及解决办法。

代码实现 #

import numpy as np
import numpy.linalg as la
#%pip install tensorly
import pandas as pd
import tensorly as tl

import warnings

from scipy.stats import norm

def matrix_outer (A, B, C ) :
    
    """
    Calculates Kronecker product of Matrices A, B, C
    """
    n_A = A.shape[0]
    
    n_B = B.shape[0]
    
    n_C = C.shape[0]
    
    k = A.shape[1]
    
    tensor = tl.zeros(shape=(n_A, n_B, n_C))
    
    for i in range(k):
        coef_a = A[:,i]
        coef_b = B[:,i]
        coef_c = C[:,i]
        tensor += tl.tenalg.tensor_dot(tl.tenalg.tensor_dot(coef_a, coef_b).reshape(n_A, n_B), coef_c).reshape(n_A, n_B, n_C)
                
    return tensor

"""
Okay now we've defined our functions.

The next step is data generation.
"""

# Tensor size

R = 4

n_user = 160 # We have 100 units/users. This is a marketing application.

n_prod = 120 # Sales data, for example, of different products

n_time = 100 # Here are our time periods.


# Auxiliary function for normalizing vectors

normalize_vec = lambda vec: vec/la.norm(vec)

"""
User participation is shown in matrix A.
"""

user_index = np.linspace(-3, 3, num=n_user)

user_bell = normalize_vec(norm.pdf(user_index, loc = 0, scale = 0.5))

user_bell2 = np.roll(user_bell,120)

user_bell3 = np.roll(user_bell,80 )

user_bell4 = np.roll(user_bell,40 )

A = np.c_[user_bell, user_bell2, user_bell3, user_bell4]

"""
Product participation is shown in matrix B.
"""

productl = normalize_vec(np.repeat([1, 2, 3 , 4], 30, axis=0))
product2 = np.roll(productl, 90 )
product3 = np.roll(productl, 60 )
product4 = np.roll(productl, 30 )

B = np.c_[productl, product2, product3, product4]


tseriesl = normalize_vec(0.3*np.sin(np.arange(0, n_time*2, step=2*np.pi/7))[0:n_time]+0.5)
tseries2 = normalize_vec(np.linspace(0, 1, n_time))
tseries3 = normalize_vec(np.repeat([0, 1], n_time/2, axis=0))
tseries4 = normalize_vec(np.repeat([1, 0], [n_time/4, 3*n_time/4], axis=0))
C = np.c_[tseriesl, tseries2, tseries3, tseries4]

#Aggregate Matrices and denoise them
np.random.seed(1512)
data = matrix_outer(A,B, C)
data_noisy= data + np.random.normal(loc=0,scale=0.5*tl.mean(data), size=(data.shape))

报错信息 #

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 14, in matrix_outer
AttributeError: module 'tensorly.tenalg' has no attribute 'tensor_dot'

原因分析 #

• 版本不匹配：通过pip安装的Tensorly稳定版可能未包含tenalg.tensor_dot——该函数是较新版本（或开发分支）中新增的，你当前安装的旧版稳定版没有这个接口。
• 导入路径差异：部分Tensorly版本中，tensor_dot函数直接放在tensorly根模块下，而非tenalg子模块内。

解决方案 #

1. 升级到最新稳定版 #

执行以下命令升级Tensorly：

pip install --upgrade tensorly

升级完成后重新运行代码，验证tl.tenalg.tensor_dot是否可用。

2. 替换为根模块的tensor_dot（兼容旧版） #

如果升级后仍无法找到该函数，可尝试将代码中的tl.tenalg.tensor_dot替换为tl.tensor_dot，旧版Tensorly的张量点积函数通常直接暴露在根模块。

3. 安装GitHub开发版（仅前两种方法无效时使用） #

若上述方案都无法解决问题，再考虑安装GitHub上的开发分支版本：

pip install git+https://github.com/tensorly/tensorly.git

代码优化建议 #

你的matrix_outer函数可以简化，直接使用张量外积函数tl.tenalg.outer代替嵌套的tensor_dot，代码更简洁高效：

def matrix_outer(A, B, C):
    """Calculates Kronecker product of Matrices A, B, C"""
    k = A.shape[1]
    tensor = tl.zeros(shape=(A.shape[0], B.shape[0], C.shape[0]))
    for i in range(k):
        # 直接计算三个向量的外积，无需多次reshape
        tensor += tl.tenalg.outer([A[:,i], B[:,i], C[:,i]])
    return tensor

内容的提问来源于stack exchange，提问作者Jared Greathouse