未定义符号：THPVariableClaload_textures.cpython37mx86_64linuxgnu.so:未定义符号：THPVariableClass

Question

你知道我怎样才能解决这个问题吗？我试图使用https://github.com/benjiebob/SMALViewer/issues/3repo，但在neural_渲染器端口上出现错误：

$ python smal_viewer.py 
Traceback (most recent call last):
  File "smal_viewer.py", line 2, in <module>
    import pyqt_viewer
  File "/home/mona/research/3danimals/SMALViewer/pyqt_viewer.py", line 13, in <module>
    from smal.smal3d_renderer import SMAL3DRenderer
  File "/home/mona/research/3danimals/SMALViewer/smal/smal3d_renderer.py", line 6, in <module>
    import neural_renderer as nr
  File "/home/mona/anaconda3/lib/python3.7/site-packages/neural_renderer/__init__.py", line 3, in <module>
    from .load_obj import load_obj
  File "/home/mona/anaconda3/lib/python3.7/site-packages/neural_renderer/load_obj.py", line 8, in <module>
    import neural_renderer.cuda.load_textures as load_textures_cuda
ImportError: /home/mona/anaconda3/lib/python3.7/site-packages/neural_renderer/cuda/load_textures.cpython-37m-x86_64-linux-gnu.so: undefined symbol: THPVariableClass

以下是一些细节：

$ python
Python 3.7.6 (default, Jan  8 2020, 19:59:22) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.__version__
'1.6.0'
>>> torch.version.cuda
'10.1'
>>> torch.cuda.is_available()
True


$ gcc --version
gcc (Ubuntu 9.3.0-10ubuntu2) 9.3.0
Copyright (C) 2019 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.


$ lsb_release -a
LSB Version:    core-11.1.0ubuntu2-noarch:security-11.1.0ubuntu2-noarch
Distributor ID: Ubuntu
Description:    Ubuntu 20.04.1 LTS
Release:    20.04
Codename:   focal

这是神经渲染器git repo:https://github.com/daniilidis-group/neural_renderer

我使用pip install neural_renderer_pytorch安装了神经渲染器

Answer 1

我首先找到所有的AT_CHECK并将它们转换为AT_ASSERT，从而解决了这个问题。检查at_assert_fix分支并没有为我解决问题。我正在使用PyTorch 1.6和CUDA 10.1

https://github.com/daniilidis-group/neural_renderer/search?q=AT_CHECK

(base) mona@mona:~/neural_renderer/neural_renderer/cuda$ vi create_texture_image_cuda.cpp 
(base) mona@mona:~/neural_renderer/neural_renderer/cuda$ vi load_textures_cuda.cpp 
(base) mona@mona:~/neural_renderer/neural_renderer/cuda$ vi rasterize_cuda.cpp 

(base) mona@mona:~/neural_renderer$ pip install .
Processing /home/mona/neural_renderer
Building wheels for collected packages: neural-renderer-pytorch
  Building wheel for neural-renderer-pytorch (setup.py) ... done
  Created wheel for neural-renderer-pytorch: filename=neural_renderer_pytorch-1.1.3-cp37-cp37m-linux_x86_64.whl size=6215781 sha256=fbd9e8ef7340fa71b9fcd69da8ee06a1079ebbe5b4c2c3ce92e8124bc8cea7c5
  Stored in directory: /tmp/pip-ephem-wheel-cache-tddilaof/wheels/c6/9b/9b/d2cda4f9ac2127278c21ea5c5e23c3354fe0e63365b7af7842
Successfully built neural-renderer-pytorch
Installing collected packages: neural-renderer-pytorch
Successfully installed neural-renderer-pytorch-1.1.3
(base) mona@mona:~/neural_renderer$ python examples/example1.py 
Drawing: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 90/90 [00:04<00:00, 19.37it/s]

https://github.com/daniilidis-group/neural_renderer/issues/94

Answer 2

看起来你不是自己造的neural_renderer_pytorch，而是用了一个轮子。但是，此控制盘是使用较旧的pytorch版本构建的，不能与计算机上当前的pytorch版本一起使用

使用当前的pytorch版本，即

$ pip uninstall neural-renderer-pytorch

$ pip install https://github.com/daniilidis-group/neural_renderer/zipball/master

它应该会起作用

---

在pytorch 1.5之前，它使用了一种在Linux上构建扩展的脆弱方式：尽管依赖于^{}，但扩展并没有显式地链接到libtorch.so。提供缺少的符号只是因为import torch用RTLD_GLOBAL加载了libtorch.so，从而使其符号全局可见/可访问-这就是为什么在加载这些扩展之前（例如neural_renderer_pytorch像here），必须导入torch的原因

在第一次导入torch之前，可以强制执行旧的行为设置RTLD_GLOBAL，以进行导入：

import sys; import ctypes;
sys.setdlopenflags(sys.getdlopenflags() | ctypes.RTLD_GLOBAL)
import torch # now all symbols of torch
             # have global visibility and can be used in 
             # other extensions

然而，使用RTLD_GLOBAL是相当危险的，因为它可能插入不相关的符号，并导致微妙的错误甚至崩溃

因此，由于1.5 pytorch不再使用RTLD_GLOBAL，而是显式地针对libpytorch.so（参见此commit）的链接和使用旧pytorch版本构建的扩展将无法工作