坑1：在 rcnn/cython 中执行 py setup.py build_ext --inplace，

出现错误 “AttributeError: 'MSVCCompiler' object has no attribute 'compiler_so'” 

修改方法：

https://github.com/deepinsight/insightface/issues/679#issuecomment-495870070   ( 感谢 loscheris )

Instead of using python implementation, you can try to compile *.pyx files into windows binaries (*.pyd) :

1. Replace setup.py in RetinaFace/rcnn/cython dir to:

   # setup.py
   from distutils.core import setup, Extension
   from Cython.Build import cythonize
   import numpy
   
   setup(
       ext_modules=cythonize(["bbox.pyx", "anchors.pyx", "cpu_nms.pyx"]),
       include_dirs=[numpy.get_include()]
   )

    

2. Use cmd to run python setup.py build_ext --inplace in RetinaFace/rcnn/cython dir

3. Copy the three windows binaries (*.pyd) generated in RetinaFace/rcnn/cython/rcnn/cpython to RetinaFace/rcnn/cython

4. Run the test code : )

坑2：ValueError: buffer dtype mismatch, expected 'int_t' but got 'long long'

解决方法：https://github.com/CharlesShang/FastMaskRCNN/issues/163#issuecomment-338623985 (感谢 mengcz13)

Fixed by changing line 25 of cpu_num.pyx from

cdef np.ndarray[np.int_t, ndim=1] order = scores.argsort()[::-1]

to

cdef np.ndarray[np.intp_t, ndim=1] order = scores.argsort()[::-1]

I am having some problems when compiling gpu_nms.pyx on Windows Server, and I am using cpu_nms. Still I don't know what will happen if you use gpu_nms...