Caffe_metal_iOS

Caffe-iOS is modifed from the master version of Caffe, a popular deep learning platform. It is designed for those inference tasks on mobile devices in iOS.

• Plug-in-and-play: General models can be directly compiled and run with corresponding parameter files. All files required are exactly what needed in general Caffe.
• Memory-optimized: Specific memory optimization for single inference tasks. Reduces 95% memory requirment on ResNet related tasks.
• GPU-accelerated: Core math computations are equipped with GPU acceleration (Metal2 Performance Shaders). Raw implementation without careful optimization has already 50% performance enhancement.

Tested Platforms

• macOS 10.13
• iOS 11.0+

Example: Neural Style Transfer on iOS

1. Open the Xcode project and setup with your own developer id and buddle.
2. Chosse "MacOS" as target and run the desktop version
3. Choose "iOS devicecs" as target and run the iOS version in iPhone, currently only iPhone 8 has been tested.
4. Choose "ios_sim" to run simulator, which may lead to some problems with insufficient simulation.

TODO:

• Complete full metal support of math functions as well as specific operations
• Provide libdnn suppport for Convolution and Deconvolution
• Implement half precision floating point calculation
• Test all metal implementation
• Optimization of metal implementation

About libdnn implementation of convolution

The metal version of libdnn implementation is borrowed from the openCL version of Caffe. The main concern of adopting libdnn is because we found that the gemm API provided by Apple may have memory leaking issue and that interface is heavily used in Convolution and Deconvolution. If the memory issue solved, we may provide another version of Convolution and Deconvolution with further speed enhancement. After adopting libdnn, we found another improvement in reduction of memeory usage.