来自 MEX 文件的意外标准异常(pytorch 模型向前) [英] Unexpected Standard exception from MEX file (pytorch model forward)
问题描述
当我从 Matlab 调用 mex api 时，出现了意外的标准异常.
When I call mex api from Matlab, I got an unexpected standard exception.
我将 2 个 pytorch DNN 模型导出到A.pt"和B.pt"文件.
I exported 2 pytorch DNN models to 'A.pt' and 'B.pt' files.
并且我实现了从.pt"文件加载模型并运行模型(向前)的 C++ 函数.
And I implemented c++ functions that load models from the '.pt 'files and run models (forward).
c++ 实现工作正常，我可以从模型中得到正确的结果.
The c++ implementation works fine, I can get proper results from the models.
我构建了负载 &将前向函数运行到 '.dll' 库中，我实现了一个可以调用它们的 mex api 函数.
I built the load & run forward function into '.dll' library, and I implemented a mex api function can call them.
当我在Matlab环境中调用mex api时，2个模型加载正常，第一个模型正常向前运行.
When I call the mex api in Matlab environment, the 2 models are loaded normally, and the first model runs forward properly.
然而，当运行第二个模型时，我得到了以下异常.
However, when run forward the 2nd model, I got the following exception.
来自 MEX 文件的意外标准异常
Unexpected Standard exception from MEX file
What():以下操作在 TorchScript 解释器中失败.
What():The following operation failed in the TorchScript interpreter.
TorchScript 的回溯(最近一次调用):
Traceback of TorchScript (most recent call last):
运行时错误:TorchScript 解释器中的以下操作失败.
RuntimeError: The following operation failed in the TorchScript interpreter.
TorchScript 的回溯(最近一次调用):
Traceback of TorchScript (most recent call last):
运行时错误:LoadLibraryA 中的错误
RuntimeError: error in LoadLibraryA
我不知道为什么 c++ 实现可以正常工作，但是当通过 Matlab 的 mex api 调用它时会发生异常.
I have no clue why the c++ implementation works fine but the exception occurs when call it through mex api from Matlab.
因为负载&run forward 函数没有改变，我期望得到完全相同的结果.
Because the load & run forward functions are unchanged I expected the exactly same results.
因为没有调用堆栈打印，所以调试起来比较困难.
It is more difficult to debug because there is no call-stack print.
有什么办法可以得到调用栈吗?
Is there any way to get call-stack ?
请给我任何建议.
提前致谢.
-环境------------------------------
-environment-------------------------------
c++ 编译器:visual studio 2017 社区
c++ compiler : visual studio 2017 community
matlab : R2020a
matlab : R2020a
libtorch : 1.6
libtorch : 1.6
pytorch : 1.5
pytorch : 1.5
蟒蛇:3.6
cuda:10.2
推荐答案
根据 Cris Luengo 先生的评论，我通过将所有 libtorch dll 复制到 Matlab 自己的 bin 文件夹中解决了这个问题.有几个重复的文件，但我覆盖了它们.我不确定它是否安全，因此备份以前的 dll 可能是不错的选择.谢谢 Cris Luengo 先生.
From Mr. Cris Luengo's comments, I solved this problem by copying all libtorch dlls into Matlab's own bin folder. There are several duplicated files but I overwrote them. I'm not sure it is safe or not, so may be backup of previous dlls is good choice. Thank you Mr. Cris Luengo.
这篇关于来自 MEX 文件的意外标准异常(pytorch 模型向前)的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
