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.

Setup step:

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.

GPU	Compute Capability
GTX660, 680, 760, 770	compute_30,sm_30
GTX780, Titan Z, Titan Black, K20, K40	compute_35,sm_35
GTX960, 980, Titan X	compute_52,sm_52

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

Matlab Wrapper

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

Python Wrapper

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.

MNIST example

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.

Update log

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.

2015/08/18 The lmdb problem has been fixed. Download the new lmdb lib file from http://pan.baidu.com/s/1dDHbbgP (only a small patch), overwrite the original one in 3rdparty/lib, and re-link the convert_imageset, convert_mnist etc projects, you will be able to create lmdb on Windows.

2015/08/08 The cuDNN v3 is not very stable at present. The master branch has been rolled back to cuDNN v2. The cuDNN v3 will come back as soon as it has been tested enough. Nonetheless, you can still find cuDNN v3 version in branch cuDNNV3.

Fortunately, cuDNN is backward-compatible, so the 3rdparty libraries (http://pan.baidu.com/s/1i390tZB) need not to be changed.

2015/08/06 cuDNN v3 is released! The new 3rdparty library with cuDNN v3 can be downloaded from http://pan.baidu.com/s/1i390tZB. In this update, I use an ungainly method to build the caffe core functions in one project as a static lib. I am still looking for better solutions. Issues and Pull Requests are welcomed.

Please help me test the speed of cuDNN v3 on non-MaxWell architecture GPUs. On my GTX780, some kinds of net, such as VGG, are quite slower than cuDNN v2.

WARNING: Visual Studio 2012 and CUDA6.5 are no longer supported. Please update your CUDA to version 7.0. If you are still using VS2012, please try this solution file and 3rdparty library http://pan.baidu.com/s/1i3hGef7. I haven't check it. So if you find bugs, please report to me.

Acknowlegement

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.