Pytorch报错(Windows平台)：from torch._C import * ImportError: DLL load failed: 找不到指定的模块。

这个错误十分奇怪，以前多次安装都没碰到过，原因可能和安装过程有关。貌似这段时间外网信号非常糟糕，我的安装过程有多次中断重启。

如何安装 pytorch请参考：https://blog.csdn.net/tanmx219/article/details/82814738

如果你的网络不好，经常需要安装本地已经下载下来的包，并且需要anaconda自动处理相关依赖，请重点参考其中的离线安装方法（2）。

下面言归正传。

报错内容

如下：

(base) C:\Users\Administrator>conda activate torch

(torch) C:\Users\Administrator>python
Python 3.7.2 (default, Feb 11 2019, 14:11:50) [MSC v.1915 64 bit (AMD64)] :: Anaconda, Inc. on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "d:\Anaconda3\envs\torch\lib\site-packages\torch\__init__.py", line 84, in <module>
    from torch._C import *
ImportError: DLL load failed: 找不到指定的模块。
>>>

原因：找不到pytorch库文件。

解决方案

Python添加库文件的方式十分简单，基本上找得到正确的库文件就不会报错，所以根据这条线，我简单地给出解决方案：

电脑==》属性==》高级系统设置 ==》高级==》环境变量==》系统 ==》Path

添加下面这一条路径：D:\Anaconda3\envs\torch\Lib\site-packages\torch\lib

如果你打开该路径就会看到，此处就是Pytorch的_C.lib等库文件的藏身之处。

 

进一步分析

（说明：后面的内容是碰到问题时的思路，目前还没有结果，先记录下  '-_-' ，有空再看）

正常情况下，anaconda是可以找到目录的，找不到的原因可能是某地库的地址记录不全或安装过程中曾经出现过错误。所以，如果pytorch是安装在anaconda的base环境中，一般也不容易产生这种问题。如果是安装在自己用conda create - n <name + option> 创建的环境中，就容易产生这种问题。

在windows下Python是如何找目录的呢？下面这一段是在其源码getpath.c中的一段注释

/* ----------------------------------------------------------------
   PATH RULES FOR WINDOWS:
   This describes how sys.path is formed on Windows.  It describes the
   functionality, not the implementation (ie, the order in which these
   are actually fetched is different). The presence of a python._pth or
   pythonXY._pth file alongside the program overrides these rules - see
   below.

   * Python always adds an empty entry at the start, which corresponds
     to the current directory.

   * If the PYTHONPATH env. var. exists, its entries are added next.

   * We look in the registry for "application paths" - that is, sub-keys
     under the main PythonPath registry key.  These are added next (the
     order of sub-key processing is undefined).
     HKEY_CURRENT_USER is searched and added first.
     HKEY_LOCAL_MACHINE is searched and added next.
     (Note that all known installers only use HKLM, so HKCU is typically
     empty)

   * We attempt to locate the "Python Home" - if the PYTHONHOME env var
     is set, we believe it.  Otherwise, we use the path of our host .EXE's
     to try and locate one of our "landmarks" and deduce our home.
     - If we DO have a Python Home: The relevant sub-directories (Lib,
       DLLs, etc) are based on the Python Home
     - If we DO NOT have a Python Home, the core Python Path is
       loaded from the registry.  This is the main PythonPath key,
       and both HKLM and HKCU are combined to form the path)

   * Iff - we can not locate the Python Home, have not had a PYTHONPATH
     specified, and can't locate any Registry entries (ie, we have _nothing_
     we can assume is a good path), a default path with relative entries is
     used (eg. .\Lib;.\DLLs, etc)

   If a '._pth' file exists adjacent to the executable with the same base name
   (e.g. python._pth adjacent to python.exe) or adjacent to the shared library
   (e.g. python36._pth adjacent to python36.dll), it is used in preference to
   the above process. The shared library file takes precedence over the
   executable. The path file must contain a list of paths to add to sys.path,
   one per line. Each path is relative to the directory containing the file.
   Blank lines and comments beginning with '#' are permitted.

   In the presence of this ._pth file, no other paths are added to the search
   path, the registry finder is not enabled, site.py is not imported and
   isolated mode is enabled. The site package can be enabled by including a
   line reading "import site"; no other imports are recognized. Any invalid
   entry (other than directories that do not exist) will result in immediate
   termination of the program.

  The end result of all this is:
  * When running python.exe, or any other .exe in the main Python directory
    (either an installed version, or directly from the PCbuild directory),
    the core path is deduced, and the core paths in the registry are
    ignored.  Other "application paths" in the registry are always read.

  * When Python is hosted in another exe (different directory, embedded via
    COM, etc), the Python Home will not be deduced, so the core path from
    the registry is used.  Other "application paths" in the registry are
    always read.

  * If Python can't find its home and there is no registry (eg, frozen
    exe, some very strange installation setup) you get a path with
    some default, but relative, paths.

  * An embedding application can use Py_SetPath() to override all of
    these automatic path computations.

  * An install of Python can fully specify the contents of sys.path using
    either a 'EXENAME._pth' or 'DLLNAME._pth' file, optionally including
    "import site" to enable the site module.

   ---------------------------------------------------------------- */

从上面的信息来看，似乎修改sys.path的路径能找到库文件，于是我在不修改环境变量的前提下，分别试了下PYTHONPATH这个环境变量，也试了一下在D:\Anaconda3\envs\torch\Lib\site-packages下面添加.path文件的办法（如下），实践证明这两者都不能让系统 找到这些_c.lib库文件，也就是说导入这些_C.lib等相关的库文件路径与sys.path这个参数无关！！！

pywin32.pth

# .pth file for the PyWin32 extensions
win32
win32\lib
Pythonwin
torch
torch\lib

如图，

Type "copyright", "credits" or "license" for more information.
IPython 7.2.0 -- An enhanced Interactive Python.
>>> import sys
>>> sys.path
Out[2]: 
['C:\\Users\\Administrator',
 'C:\\Users\\Administrator',
 'D:\\Anaconda3\\envs\\torch\\Lib\\site-packages\\torch\\lib',
 'D:\\Anaconda3\\envs\\torch\\python37.zip',
 'D:\\Anaconda3\\envs\\torch\\DLLs',
 'D:\\Anaconda3\\envs\\torch\\lib',
 'D:\\Anaconda3\\envs\\torch',
 '',
 'D:\\Anaconda3\\envs\\torch\\lib\\site-packages',
 'D:\\Anaconda3\\envs\\torch\\lib\\site-packages\\win32',
 'D:\\Anaconda3\\envs\\torch\\lib\\site-packages\\win32\\lib',
 'D:\\Anaconda3\\envs\\torch\\lib\\site-packages\\Pythonwin',
 'D:\\Anaconda3\\envs\\torch\\lib\\site-packages\\torch',
 'D:\\Anaconda3\\envs\\torch\\lib\\site-packages\\IPython\\extensions',
 'C:\\Users\\Administrator\\.ipython']

>>> import torch
Traceback (most recent call last):

  File "<ipython-input-4-eb42ca6e4af3>", line 1, in <module>
    import torch

  File "D:\Anaconda3\envs\torch\lib\site-packages\torch\__init__.py", line 84, in <module>
    from torch._C import *

ImportError: DLL load failed: 找不到指定的模块。

 

一些参考：

1。pytorch开发团队认为，这种问题是因为在当前目前下也有一个torch文件夹，这个文件夹被误认为是torch库所在的文件夹所致，可惜我的问题不属于该种情况，

https://github.com/pytorch/pytorch/issues/574
the problem is that you have a folder called `torch` in the same directory which is being picked up. Do this: `cd ..` (to change directory), and then start `python` and `import torch`, it should work.

2。pip install numpy -I
这个办法来自https://github.com/pytorch/pytorch/issues/2731

只能说这是个非常糟糕的办法，我使用该命令后陷入了一个更大的困境，如下，

Traceback (most recent call last):
File "", line 1, in 
File "/home/user/anaconda2/envs/py35/lib/python3.5/site-packages/torch/**init**.py", line 53, in 
from torch._C import *
ImportError: numpy.core.multiarray failed to import