Awesome
Windows Caffe
This is an experimental, I would migrate ssd from Linux to Windows, and based on @willyd work. It is a work-in-progress.
This branch of Caffe ports the framework to Windows.
Update: this branch is not actively maintained. Please checkout this for more active Windows support. If you want to read the Chines version of READMER, please click on it please click README-Chinese
Windows Setup
Requirements
Optional Dependencies
- Python for the pycaffe interface. Anaconda Python 2.7 or 3.5 x64 (or Miniconda)
- Matlab for the matcaffe interface.
- CUDA 7.5 or 8.0 (use CUDA 8 if using Visual Studio 2015)
- cuDNN v5
We assume that cmake.exe
and python.exe
are on your PATH
.
Configuring and Building Caffe
The fastest method to get started with caffe on Windows is by executing the following commands in a cmd
prompt (we use C:\Projects
as a root folder for the remainder of the instructions):
C:\Projects> https://github.com/runhang/caffe-ssd.git
C:\Projects> cd caffe-ssd
:: Edit any of the options inside build_win.cmd to suit your needs
C:\Projects\caffe> scripts\build_win.cmd
The build_win.cmd
script will download the dependencies, create the Visual Studio project files (or the ninja build files) and build the Release configuration. By default all the required DLLs will be copied (or hard linked when possible) next to the consuming binaries. If you wish to disable this option, you can by changing the command line option -DCOPY_PREREQUISITES=0
. The prebuilt libraries also provide a prependpath.bat
batch script that can temporarily modify your PATH
envrionment variable to make the required DLLs available.
Below is a more complete description of some of the steps involved in building caffe.
Install the caffe dependencies
By default CMake will download and extract prebuilt dependencies for your compiler and python version. It will create a folder called libraries
containing all the required dependencies inside your build folder. Alternatively you can build them yourself by following the instructions in the caffe-builder README.
Use cuDNN
To use cuDNN the easiest way is to copy the content of the cuda
folder into your CUDA toolkit installation directory. For example if you installed CUDA 8.0 and downloaded cudnn-8.0-windows10-x64-v5.1.zip you should copy the content of the cuda
directory to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0
. Alternatively, you can define the CUDNN_ROOT cache variable to point to where you unpacked the cuDNN files e.g. C:/Projects/caffe/cudnn-8.0-windows10-x64-v5.1/cuda
. For example the command in scripts/build_win.cmd would become:
cmake -G"!CMAKE_GENERATOR!" ^
-DBLAS=Open ^
-DCMAKE_BUILD_TYPE:STRING=%CMAKE_CONFIG% ^
-DBUILD_SHARED_LIBS:BOOL=%CMAKE_BUILD_SHARED_LIBS% ^
-DBUILD_python:BOOL=%BUILD_PYTHON% ^
-DBUILD_python_layer:BOOL=%BUILD_PYTHON_LAYER% ^
-DBUILD_matlab:BOOL=%BUILD_MATLAB% ^
-DCPU_ONLY:BOOL=%CPU_ONLY% ^
-DCUDNN_ROOT=C:/Projects/caffe/cudnn-8.0-windows10-x64-v5.1/cuda ^
-C "%cd%\libraries\caffe-builder-config.cmake" ^
"%~dp0\.."
Alternatively, you can open cmake-gui.exe
and set the variable from there and click Generate
.
Building only for CPU
If CUDA is not installed Caffe will default to a CPU_ONLY build. If you have CUDA installed but want a CPU only build you may use the CMake option -DCPU_ONLY=1
.
Using the Python interface
The recommended Python distribution is Anaconda or Miniconda. To successfully build the python interface you need to add the following conda channels:
conda config --add channels conda-forge
conda config --add channels willyd
and install the following packages:
conda install --yes cmake ninja numpy scipy protobuf==3.1.0 six scikit-image pyyaml pydotplus graphviz
If Python is installed the default is to build the python interface and python layers. If you wish to disable the python layers or the python build use the CMake options -DBUILD_python_layer=0
and -DBUILD_python=0
respectively. In order to use the python interface you need to either add the C:\Projects\caffe\python
folder to your python path of copy the C:\Projects\caffe\python\caffe
folder to your site_packages
folder.
Using the MATLAB interface
Follow the above procedure and use -DBUILD_matlab=ON
. Change your current directory in MATLAB to C:\Projects\caffe\matlab
and run the following command to run the tests:
>> caffe.run_tests()
If all tests pass you can test if the classification_demo works as well. First, from C:\Projects\caffe
run python scripts\download_model_binary.py models\bvlc_reference_caffenet
to download the pre-trained caffemodel from the model zoo. Then change your MATLAB directory to C:\Projects\caffe\matlab\demo
and run classification_demo
.
Using the Ninja generator
You can choose to use the Ninja generator instead of Visual Studio for faster builds. To do so, change the option set WITH_NINJA=1
in the build_win.cmd
script. To install Ninja you can download the executable from github or install it via conda:
> conda config --add channels conda-forge
> conda install ninja --yes
When working with ninja you don't have the Visual Studio solutions as ninja is more akin to make. An alternative is to use Visual Studio Code with the CMake extensions and C++ extensions.
Building a shared library
CMake can be used to build a shared library instead of the default static library. To do so follow the above procedure and use -DBUILD_SHARED_LIBS=ON
. Please note however, that some tests (more specifically the solver related tests) will fail since both the test exectuable and caffe library do not share static objects contained in the protobuf library.
Troubleshooting
Should you encounter any error please post the output of the above commands by redirecting the output to a file and open a topic on the caffe-users list mailing list.
Known issues
- The
GPUTimer
related test cases always fail on Windows. This seems to be a difference between UNIX and Windows. - Shared library (DLL) build will have failing tests.
- Shared library build only works with the Ninja generator
Further Details
Refer to the BVLC/caffe master branch README for all other details such as license, citation, and so on.