Faster-rcnn-matlab-cuda8.0+zfnet Training Your own data

The main reference to this blog post, step by step configuration can run up ~ ~

First of all the hardware you need: I started with GT630, only 2G memory, the program ran to half of the error out of memory to know at least 3G video memory to train ZF net, as for VGG-16 Net is required up to 8G of video memory. So, change the machine of the laboratory, with GTX1080 try.

As for the software, win7 system + MATLAB 2014b + Cuda 8.0 + vs 2013, nothing else to say, Cuda 8.0 is 1080 standard, no no No. With cuda8.0 's sake, the Mex file needs to compile itself, this is a giant pit, but it is best to learn to compile their own, the later version of the province has changed its own to go online to find someone else to compile the finished product.

Let's start with the vs2013+cuda8.0 compiled Caffe external file ~ (VS and Cuda installation is very brain-free, not to speak, it is recommended to first install vs again cuda)

Compile process reference here: Caffe for Faster r-cnn, just follow the steps.

It looks simple, doesn't it? Operation up a lot of pits!!

1) Download Caffe_library

2) Copy the Caffe on GitHub to the Caffe directory in Caffe_library

3) Prepare OPENCV,BOOST,MKL

A) Boost no matter what version you must pay attention to download x64-bit installation package, after installation will have lib64-msvc-12.0 such a folder will be used later.

Online said must be compiled to use, but in fact lib64-msvc-12.0 has precompiled all of the libraries we need to use directly.

b) OpenCV use of the non-compiled package;

c) MKL needs to be installed

4) Open the SLU in the Caffe_library directory and change the configuration to Release_mex

Open when found that the project Caffe can not be loaded, it is because the default is cuda6.5, so to right-click Caffe Project, modify the configuration file Cuda version number

5) Modify the include path and the Lib path

How to operate can Baidu, the path exactly to which sub-folder can refer to the author (the author has added his own computer on the external library path, you only need to do a little modification can be)

6) Set Caffe as Startup Project

7) Recompile the Caffe package

There is also a pit, because boost has been used c++14 standard, and VS2013 to update5 only support c++14, so there is no way to only obediently upgrade to Update5.

8) Copy all the files in the. \x64\release_mex to the Faster_rcnn-master\external \caffe\matlab\caffe_faster_rcnn.

I have compiled a library here, if the same as my environment can copy the past

So far, FASTER-RCNN's external library has been updated ~~~~~~

Sprinkle flower ~ ~ ~

Then run FASTER_RCNN-MASTER\FASTER_RCNN_BUILD.M not expected, error!

At this time find the. \function\nms\nvmex.m file, modify host_copliler_location for your computer vs2012 path, important thing to say three times!

! 2012! 2012! 2012!

That is to say you want to install vs2012, anyway my 2013 not, do not know there are other ways.

Compiling the Mex file means the preparatory work is complete.

Next to prepare your own data set, you can refer to this blog, the use of the tool can be downloaded from here.

The pro-Test labeling tool can only be used under Win7, WIN8, win10 are not.

Follow tutorial One by one to modify the file , run EXPERIMENTS/SCRIPT_FASTER_RCNN_VOC2007_ZF.M will soon be able to enter the iteration.

Now my 1080 is running at a high speed and I think it will be possible tomorrow morning.

Faster-rcnn-matlab-cuda8.0+zfnet Training Your own data