Win10 + Python + MXNet + VS2015 configuration

Projects need to use MTCNN to detect, align, cut out the face, it is using mxnet as a framework, but my own Ubuntu in various frameworks mess, do not want to mess up the heart to be in Windows with a. Helpless online information is not much, struggled a few days later decided to leave such a document.

The first thing we're using is not the set of mxnet that DMLC on GitHub, which is not updated in the Windows edition of 2016, and now the address for maintenance is this:

Https://github.com/yajiedesign/mxnet/releases

This is a more daily version, very new. Coincidentally is MTCNN also use the new version of the mxnet, so the 16 version of the bad so that, rely on this day more, the following about this set of products collectively referred to as "the Japanese version."

There are some parameters in the new version of the mxnet in the VC12 compiled library, so we recommend the use of VC14 version, that is, VS2015.

Before formally installing the configuration steps, we recommend a software called Dependency Walker, whose role is to clarify the dependencies of the library files and to help you later to rely on the version of the library:

This is the use of software, note the Red box section, here we need to be a bit sensitive to the file name:

1, Libopenblas Nothing special, the Japanese version of the inside with.

2, cudart64_80, curand64_80, cublas64_80, cufft64_80, nvrtc64_80 these four pay attention to the following numbers, this group of files can be found in 3rdparty\cudart, if the version number is not the same, Please make your own corrections and the daily version should bring your own.

3, Cudnn64_5 This is the need to download the file, from the file name can be known we need a 64-bit CUDNN, the version number is about 5 beginning, according to the above statement we also need a CUDA8.0 with the CUDNN.

4, Nvcuda This is obviously Nvidia Cuda, but no version number first regardless of it.

5, VCOMP140 This can be found in 3RDPARTY\VC, the version number is not correct, please modify

6, KERNEL32, USER32 this kind of public face should not be any special demand, whether they

[Winerror 126], which appears during import, is basically for this reason.

———————————————————————— the next step is to begin formally ————————————————————————————

1. Download and unzip the VC14 base package in the daily version address

2. Download and unzip the 2017xxxx_mxnet_x64_vc14_gpu.7z in the address of the daily version into the directory of the base package above to form a complete structure

2, one of the most important things is the other various tutorials that need make ah, compile ah of the Build\libmxnet.dll, that is, the above dependency Walker concerned about this document, So the precompiled version is the big guy helped us compile this DLL beforehand.

3, [Winerror 126] appears because the import of a module, it depends on the library needs to be able to find the library must be found, or error.

4, according to the above said installed good cuda8.0, this tutorial a lot of also very good outfit.

5. Ready CUDNN5 Series for cuda8.0

6, focus on the Japanese version of the file, first run the root directory of the Setupenv.cmd, this file is set all the relevant environment variables

It is recommended to take a look at the path in the "User variables" before you start, and I have a problem with the truncation of characters in the process, in other words, my previous user variables have been partially destroyed, so be prepared ahead of time.

7. command line to Python directory, execute

Install

8, finally into the mxnet\3rdparty to the cudnn of the relevant documents copied into it, and other dependent files should be in

9, you can enter Python to try import mxnet.

I wish you success.

Win10 + Python + MXNet + VS2015 configuration