windows环境下编译warp-ctc

最近在学习文本检测与识别

此次博文是一次失败的安装过程，最后用了pytorch中自带的CTCLoss，因为我的pytorch为1.1.0版本

如果版本为1.0以下，可以尝试以下方法，本人已经走到了最后一步，成功安装，但是却无法调用。

主要过程如下：

1.安装cmake；

2.下载warp-ctc，下载链接：https://github.com/amberblade/warp-ctc

以上链接是作者修改之后，可以在windows环境下编译

3.在命令行，进入warp-ctc，并建立文件目录build，进入该目录，执行cmake -G "Visual Studio 14 2015 Win64" ..

我需要到了一个错误，找不到vs编译器，后来查找了资料，在vs2015中，将SDK10添加安装后，编译成功。

到此编译成功，但是打不开sln文件，计算机重启后，重新编译了一次，打开sln，又需要安装适用于 Windows 桌面的 Visual C++ 2015 工具SDK8.1，安装之后，编译ALL_BUILD成功。

编译成功后，修改setup.py为以下内容：

# build.py
import os
import platform
import sys
from setuptools import setup, find_packages

#from torch.utils.ffi import create_extension
from torch.utils.cpp_extension import BuildExtension
from torch.utils.cpp_extension import CppExtension
import torch


extra_compile_args = ['-std=c++11', '-fPIC', '-std=c99']
warp_ctc_path = "../build/release"

if torch.cuda.is_available() or "CUDA_HOME" in os.environ:
    enable_gpu = True
else:
    print("Torch was not built with CUDA support, not building warp-ctc GPU extensions.")
    enable_gpu = False

lib_ext = ".dll"

headers = ['src/cpu_binding.h']

# if enable_gpu:
#     extra_compile_args += ['-DWARPCTC_ENABLE_GPU']
#     headers += ['src/gpu_binding.h']

if enable_gpu:
    from torch.utils.cpp_extension import CUDAExtension

    build_extension = CUDAExtension
    extra_compile_args += ['-DWARPCTC_ENABLE_GPU']
else:
    build_extension = CppExtension

if not os.path.exists(os.path.join(warp_ctc_path, "warpctc" + lib_ext)):
    print(("Could not find libwarpctc.so in {}.\n"
           "Build warp-ctc and set WARP_CTC_PATH to the location of"
           " warpctc.dll (default is '../build/release')").format(warp_ctc_path))
    sys.exit(1)
include_dirs = [os.path.realpath('../include')]

# ffi = BuildExtension(
#     name='warpctc_pytorch._warp_ctc',
#     package=True,
#     language='c++',
#     headers=headers,
#     sources=['src/binding.cpp'],
#     with_cuda=enable_gpu,
#     include_dirs=include_dirs,
#     library_dirs=[os.path.realpath(warp_ctc_path)],
#     libraries=['warpctc', 'caffe2_gpu', '_c', 'caffe2'],
#     extra_link_args=['-Wl,-rpath,' + os.path.realpath(warp_ctc_path)],
#     extra_compile_args=extra_compile_args)
# ffi = ffi.distutils_extension()
# setup(
#     name="warpctc_pytorch",
#     version="0.1",
#     description="PyTorch wrapper for warp-ctc",
#     url="https://github.com/baidu-research/warp-ctc",
#     author="Jared Casper, Sean Naren",
#     author_email="jared.casper@baidu.com, sean.narenthiran@digitalreasoning.com",
#     license="Apache",
#     packages=find_packages(),
#     ext_modules=[ffi],
# )
ext_modules = [
    build_extension(
        name='warpctc_pytorch._warp_ctc',
        package=True,
        language='c++',
        headers=headers,
        sources=['src/binding.cpp'],
        include_dirs=include_dirs,
        library_dirs=[os.path.realpath(warp_ctc_path)],
        libraries=['warpctc', 'caffe2_gpu', '_c', 'caffe2'],
        #libraries=['warpctc'],
        extra_link_args=['-Wl,-rpath,' + os.path.realpath(warp_ctc_path)],
        extra_compile_args=extra_compile_args
    )
]

setup(
    name="warpctc_pytorch",
    version="0.1",
    description="PyTorch wrapper for warp-ctc",
    url="https://github.com/baidu-research/warp-ctc",
    author="Jared Casper, Sean Naren",
    author_email="jared.casper@baidu.com, sean.narenthiran@digitalreasoning.com",
    license="Apache",
    packages=find_packages(),
    ext_modules=ext_modules,
    cmdclass={'build_ext': BuildExtension}
)

在命令行输入：python setup.py install

安装成功后，将dll文件复制到安装目录下。

但是测试时仍然保存。不知道什么原因。

最后还是使用了pytorch自带的ctcloss。

参考：

https://blog.csdn.net/qq_41895190/article/details/82818615

https://blog.csdn.net/fanhenghui/article/details/102745694