windows下pytorch版faster rcnn代码使用

本文主要是介绍windows下pytorch版faster rcnn代码使用，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

faster rcnn_pytorch代码使用

使用jwyang的代码，但是原github上是linux环境下的。这里尝试windows下的使用。

compile modules问题调试

安装步骤中

Compile modules(nms, roi_pooling, and roi_align(from facebookresearch/maskrcnn-benchmark)):

cd pytorch-faster-rcnn/lib
python setup.py build develop
cd ../

出现的问题和解决方式如下：

1.unsupported Microsoft Visual Studio version

Error C1189 #error: – unsupported Microsoft Visual Studio version! Only the versions 2012, 2013, 2015 and 2017 are supported!

参考web

With some debugging, I found that on line 131 of file host_config.h in directory “C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\include\crt” only supports a max _MSC_VER of 1910. Since Visual Studio 2017 is reporting version 1911, I changed the line to:

#if _MSC_VER < 1600 || _MSC_VER > 1911

And now the new project will build and run.

2.windows查看cuda和cudnn版本

CUDA：powershell输入nvcc -V

CUDNN：cudnn.h中MAJOR和MINOR

3.error: no instance of function template “THCCeilDiv” matches the argument list

error: no instance of overloaded function “std::min” matches the argument list

参考issue254和danpe1327

在出现错误的代码中修改为

define a function ceil_div which is

int ceil_div(int a, int b):return (a + b - 1) / b;

replace the long with int

dim3 grid(std::min(ceil_div((int)grad.numel(), 512), 4096));

注意danpe1327仅作参考修改ROIPool_cuda.cu和ROIAlign_cuda.cu，不需要按照其中内容修改setup.py。

4.error: could not create ‘maskrcnn_benchmark_C.cp36-win_amd64.pyd’: No such file or directory

如果修改了setup.py，就会出现这个问题。

Install the Python COCO API.

Install the Python COCO API. The code requires the API to access COCO dataset.

cd data
git clone https://github.com/pdollar/coco.git
cd coco/PythonAPI
make
cd ../../..

window时中安装cocoapi使用以下命令：

pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI

参考win_cocoapi

问题：ModuleNotFoundError: No module named ‘pycocotools._mask’

If you get this error, it is because the COCO library you are calling is NOT our version of the library. Instead, you are (inadvertently) referencing a local copy of the library. So, if you get this error, search the tree of import in your code for a reference to a local pycocotools library. Then, move or delete the local version.

当调试trainval.py时，模块调用不出现error说明以上配置完成。

训练

将预训练模型放置在/data/pretrained_model上；

将数据文件放置在/data中，如VOC数据集放置目录如下：

|-data
|   |-VOCdevkit2007
|           |-VOC2007
|                |-Annotations
|                |-ImageSets
|                |-JPEGImages
|                |-SegmentationClass
|                |-SegmentationObject

点击训练即可。

未完待续

这篇关于windows下pytorch版faster rcnn代码使用的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！