Forked from https://www.github.com/BVLC/caffe master branch in 2015/9/1

Added Batch Normalization, Parametric ReLU, Locally Connected Layer, Normalize Layer, Randomized ReLU, Triplet Loss, SmoothL1 Layer, ROI Layer.

1. Download third-party libraries from BaiduYun Disk or OneDrive and extract the files to caffe-windows_root/3rdparty/. Please don't forget to add the ./3rdparty/bin folder to your environment variable PATH.

2. Run ./src/caffe/proto/extract_proto.bat to create caffe.pb.h, caffe.pb.cc and caffe_pb2.py.

3. Double click ./build/MainBuilder.sln to open the solution.

4. Change the compile mode to Release and X64. For Debug mode, you may need these 3rparty libraries http://pan.baidu.com/s/1qW88MTY .

5. Modify the cuda device compute capability defined in the settings (caffelib properties -> CUDA C/C++ -> Device -> Code Generation) to your GPU's compute capability (such as compute_30,sm_30; etc). You can look up for your GPU's compute capability in https://en.wikipedia.org/wiki/CUDA . Some general GPUs' compute capabilities are listed below.

• If your GPU's compute capability is below or equal to 2.1, please remove the USE_CUDNN macro in the proprocessor definition of all projects.

• If you do not have a Nvidia GPU, please also add CPU_ONLY macro besides removing USE_CUDNN.

6. Compile.

TIPS: If you have MKL library, please add the preprocess macro "USE_MKL" defined in the setting of the project.

If you want build other tools, just copy and rename ./build/MSVC folder to another one, and add the new project to the VS solution. Remove caffe.cpp and add your target cpp file. Compile it, then you will get a corresponding exe file in ./bin.

中文安装说明：http://blog.csdn.net/happynear/article/details/45372231

Just replace the Matlab include and library path defined in the settings and compile. Don't forget to add ./matlab to your Matlab path.

Similar with Matlab, replace the python include and library path and compile.

Most of the libraries listed in ./python/requirements.txt can be installed by pip install. However, some of them cannot be installed so easily.

For protobuf, you may download the codes from https://github.com/google/protobuf. Copy caffe-windows-root/src/caffe/proto/protoc.exe to protobuf-root/src. Then run python setup.py install in protobuf-root/python.

For leveldb, I have created a repository https://github.com/happynear/py-leveldb-windows . Please follow the instructions in README.md to install it.

Please download the mnist leveldb database from http://pan.baidu.com/s/1mgl9ndu and extract it to ./examples/mnist. Then double click ./run_mnist.bat to run the MNIST demo.

2015/11/09 CuDNN v3 works well now.

2015/09/14 Multi-GPU is supported now.

WARNING: When you are using multiple gpus to train a model, please do not directly close the command window. Instead, please use Ctrl+C to avoid the gpu driver from crash.

You can also press Ctrl+Break to save a model snapshot whenever you want during training.

We greatly thank Yangqing Jia and BVLC group for developing Caffe,

@niuzhiheng for his contribution on the first generation of caffe-windows,

@ChenglongChen for his implementation of Batch Normalization,

@jackculpepper for his implementation of locally-connected layer,

and all people who have contributed to the caffe user group.