一定概率会出现内存会一直涨

[leoxxxxxD]

每次运行可能会出现内存泄露？现象就是内存不断占用，但是一直没有前向

[18242360613]

出现了相同的问题，问题出现在load模型的时候。

[yasenh]

@xw0629 @18242360613 thanks for the feedback! But I cannot replicate the issue locally, could you provide your testing environment (OS, CUDA version, libtorch version)?

[liubamboo]

I have the same problem. my environment is ubuntu 18.04, cuda 10.2 libtorch 1.8.0

the bug is caused by the module_ = torch::jit::load(model_path); But I don't know how to solve it.

[liubamboo]

envs: ubuntu 18.04, libtorch 1.8.1, cxxopts

Hello, When I run the code from https://github.com/yasenh/libtorch-yolov5.git. the program memery is still increasing. After debugging, I find the key is that cxxopts conflict with libtorch.

when I use cxxopts in my main.cpp and load torch script by torch::jit::load, the bug sometime appears. The program memory is still increasing quickly and finally the program was killed by the operater system, and then there are some infomation like: Process finished with exit code 9

Then I didn't use cxxopts and use gflags, The bug disappears.

So I think there are some conflicts between libtorch and cxxopts

[monoloxo]

@liubamboo 您好，我也遇到了同样的问题，我也是定位到loading model那里容易出现：killed。然后我按你的方法，把 std::string weights = opt["weights"].as<std::string>(); 改成 std::string weights = "../weights/best.torchscript-gpu-half.pt"; 但仍然有几率出现错误。。 难道是要把所有用到cxxopts的代码都替换掉吗？

[liubamboo]

@monoloxo 我使用glog替换掉cxxopts，就没有问题了，替换也比较简单，安装glog然后更改主函数开头输入参数部分即可。详细原因我还没有找到

[monoloxo]

@liubamboo 感谢，我试一下

[monoloxo]

@liubamboo 打扰一下，我刚安装之后试着写了一下，出现了好多错误。。发现要改好多东西，（刚学C++,很多不太会的），方便的话可以看一下您使用的方法吗？

[monoloxo]

感谢 @liubamboo ，确实是cxxopts的问题，我把相关的参数都改成直接输入的方式，问题就消失了。

[yasenh]

@liubamboo @monoloxo feel free to create a PR to fix it.