- It is copied from https://github.com/BVLC/caffe/tree/windows at 2017.03.13
-
Resize layer
- Layer type: "Resize"
- Resize 2D layer to a specific size
- Parameters: height, width, interpolation(method: NN, LINEAR)
- Width and height can be specified by placing a second bottom. If the second bottom is placed, the parameter input is ignored
-
Unpooling layer
- Layer type: "Unpooling"
- It is referenced from https://github.com/HyeonwooNoh/caffe
- Unpool-operation
- Parameters: pad, kernel_size, stride, unpool_size, unpool(method: MAX, AVE, TILE)
-
Batch renormalize layer
- Layer type: "BatchReNorm"
- Implementation of https://arxiv.org/abs/1702.03275
- Renormalize batch-normalization to follow global-normalization
- Parameters: use_global_stats, moving_average_fraction, eps, iter_size, step_to_init, step_to_r__max, step_to_d_max, r_max, d_max
- For training, lr_murt and decay_mult are automatically set to zero
-
Mean squared relative loss layer
- Layer type: "MeanSquaredRelativeLoss"
- Compute mean squared relative error loss from two inputs
- Loss=E(((Y-T)/T)^2)
- Parameters: ignore_value_min, ignore_value_max
- Elements with value between ignore_value_min and ignore_value_max are excluded from the loss calculation
-
Add ignore-range feature to EuclideanLossLayer
- Additional parameters: ignore_value_min, ignore_value_max
- Elements with value between ignore_value_min and ignore_value_max are excluded from the loss calculation
-
Adaptive learning rate control
- lr_policy: "adaptive"
- Adjust learning rate using relative decrement of train loss
- RelativeLossDecrement=(PreTermLossAverage-CurrentTermLossAverage)/PreTermLossAverage
- It is recorded in solver state
- Parameters: lr_control_size, initial_learning_term, max_loss, lr_decrease_factor, lr_increase_factor, min_loss_diff, max_loss_diff, reset_controled_lr
This is an experimental, communtity based branch led by Guillaume Dumont (@willyd). It is a work-in-progress.
This branch of Caffe ports the framework to Windows.
AppVeyor (Windows build)
Prebuilt binaries can be downloaded from the latest CI build on appveyor for the following configurations:
-
Visual Studio 2015, CPU only, Python 3.5: Caffe Release,
Caffe Debug -
Visual Studio 2015, CUDA 8.0, Python 3.5: Caffe Release
-
Visual Studio 2015, CPU only, Python 2.7: Caffe Release, Caffe Debug
-
Visual Studio 2015,CUDA 8.0, Python 2.7: Caffe Release
-
Visual Studio 2013, CPU only, Python 2.7: Caffe Release, Caffe Debug
- 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.
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> git clone https://github.com/BVLC/caffe.git
C:\Projects> cd caffe
C:\Projects\caffe> git checkout windows
:: Edit any of the options inside build_win.cmd to suit your needs
C:\Projects\caffe> scripts\build_win.cmdThe 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.
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.
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.
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.
The recommended Python distribution is Anaconda or Miniconda. To successfully build the python interface you need to install the following packages:
conda install --yes numpy scipy matplotlib scikit-image pip six
also you will need a protobuf python package that is compatible with pre-built dependencies. This package can be installed this way:
conda install --yes --channel willyd protobuf==3.1.0
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.
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.
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 --yesWhen 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.
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.
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.
The previous windows build based on Visual Studio project files is now deprecated. However, it is still available in the windows folder. Please see the README.md in there for details.
- The
GPUTimerrelated 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
Refer to the BVLC/caffe master branch README for all other details such as license, citation, and so on.