Embedded Caffe

  Make the depth learning model running on Embedded Systems and multi-platform.

Features

• Based on caffe
• Compact，Only forward calculation
• Support faster rcnn
• Support ssd
• Support ShuffleNet
• Support MobileNet
• Support GPU/CPU model
  • Window only support x86_64 archtecture in GPU mode.
  • CPU_ONLY mode support x86/x86_64 archtecture on Win platform.
• Cross-platform
• Small
• Remove the gtest
• Remove the test
• Remove the python wrapper

Requirements

Linux

System              : Linux / Ubuntu 14.04
Cmake               : 3.4+
C++ compiler        : 4.8+
BLAS                : Atlas
Boost               : 1.54
glog                : Y
gflags              : Y
protobuf            : 2.6.1
CUDA                : 7.5
cuDNN               : 5.0.5
OpenCV              : N (for examples)

Windows

• CPU_ONLY mode support x86/x86_64 archtecture.
• GPU model only support x86_64 archtecture, because NVIDIA only support the x86_64 archtecture libraries since CUDA7.0.

System              : Windows
Cmake-gui           : 3.8+
C++ compiler        : vc140(VS2015)
embcaffe_3rdparty   : https://github.com/FreeApe/embcaffe_3rdparty

Build

Build EmbCaffe on Linux

• Way 1

# Note: Test examples with this way
$ cd EmbCaffe
$ mkdir cmake_build
$ cd cmake_build
$ cmake ..
$ make all -j

• Way 2

$ cd EmbCaffe
$ make all -j

Build EmbCaffe on Windows

1. with CMake-gui3.8+
2. git clone https://github.com/FreeApe/embcaffe_3rdparty(You can also compile these third-party libraries yourself)
2. Configure and Generate(You should configure CMAKE_INCLUDE_PATH and CMAKE_LIBRARY_PATH)
3. Open Caffe.sln with VS2015
4. build solutions

Build examples

  Linux:

$ cd examples
$ mkdir build
$ cd build
$ cmake ..
$ make -j

  Windows:

1. with CMake-gui3.8+
2. Configure and Generate
3. Open Caffe-example.sln with VS2015
4. build solutions
5. run demos

  Running the examples, the results show as :

Build Errors

1. error ‘type name’ declared as function returning an array escape

    Make sure the CUDA version is 7.5

TODO / Targets

• Remove Backward calculation
• Support GPU model on windows
• Optimize the calculation
• ......

Thanks

• https://github.com/BVLC/caffe
• https://github.com/intel/caffe
• http://blog.csdn.net/zxj942405301/article/details/72775463