Caffe has lots of options and any small mistake can lead you off track for hours. I've successfully compiled and run Caffe for my own Macbook Pro (Retina, Mid 2012) under following conditions:
- macOS Sierra 10.12.4
- NVIDIA® CUDA™ 8.0.83 with cudnn
- Xcode CLT(Command Line Tools) 8.2 (Apple clang 80000)
- OpenCV 3.2.0
- Python 3.6
- numpy 1.12.1
- ATLAS BLAS
- boost 1.64.0
Sometimes CUDA is not supported with the latest version of the Apple LLVM and output an error message as follows.
nvcc fatal : The version ('80100') of the host compiler ('Apple clang') is not supported
The current CUDA driver 8.0.83 supports Xcode CLT 8.2 so you should switch to the old version.
- Log in to https://developer.apple.com/downloads/
- Download Xcode CLT (Command Line Tools) 8.2
- Install CLT
- Run
sudo xcode-select --switch /Library/Developer/CommandLineTools
- Verify that clang has been downgraded via clang --version