This project provides an optimized Caffe library for iOS and Android, focusing on reduced size and improved speed for mobile deployment. It targets developers building on-device machine learning applications who need a lightweight Caffe backend.

How It Works

The library is a fork of the original Caffe, heavily modified for mobile environments. Key optimizations include disabling features like HDF5, LevelDB, Boost, and backward computation, resulting in a smaller footprint and faster inference. It utilizes CMake for building and includes JNI code for Android integration.

Quick Start & Requirements

• iOS: Clone the repository, run ./tools/build_ios.sh, and then build the CaffeSimple Xcode project. Requires a pre-trained LeNet model.
• Android: Clone the repository, set NDK_HOME, and run ./tools/build_android.sh. Requires Android NDK r15 or earlier. Build the CaffeSimple Android Studio project.
• macOS/Ubuntu: Install protobuf, libprotobuf-dev, protobuf-compiler, libatlas-dev (Ubuntu) or gflags (macOS). Build with CMake.
• Windows: Requires Git for Windows, TDM-GCC, and CMake 3.8.2. Build via Git Bash.

Highlighted Details

• Optimized for size and speed by disabling non-essential Caffe features.
• Includes out-of-the-box demo applications for iOS and Android.
• Supports conversion of .prototxt files to .protobin for mobile loading.
• Build scripts leverage CMake for cross-platform compilation.

Maintenance & Community

This project appears to be a personal fork with limited recent activity. There are no links to community channels or roadmaps provided in the README.

Licensing & Compatibility

The project is based on Caffe, which is typically under a BSD-style license. However, the specific licensing for this fork is not explicitly stated in the README. Compatibility with commercial or closed-source applications would require verification of the license.

Limitations & Caveats

NDK versions r16 and later are not supported due to compatibility issues with header changes. The build process for Android requires a specific NDK version (r15 or earlier). The project's maintenance status and community support are unclear.