Mobile MN

This demo provides CMake build scripts to compile a source code, demo.cpp, that trains a LeNet model in MNN framework, into an executable target file demo.out along with MNN shared libraries for Linux/macOS and Android platforms.

Environmental Requirements

• cmake (version >=3.10 is recommended)
• protobuf (version >= 3.0 is required)
• gcc (version >= 4.9 is required)

Linux/macOS

1. run ./build_x86_linux.sh, which will generate the executable file demo.out under build_x86_linux folder.

cd build/train
sh build_x86_linux.sh
./build_x86_linux/demo.out mnist ../../../data/lenet_mnist.mnn ../../../data/mnist

# Use high precision model (Customized Softmax Operator):
./build_x86_linux/demo.out mnist ../../../data/lenet_mnist_high_precision.mnn ../../../data/mnist

# Use tabular type dataset for functional testing:
./build_x86_linux/demo.out tabular ../../../data/lenet_mnist.mnn ../../../data/tabular

Android

1. Download and Install NDK, latest release version is recommended
2. Set ANDROID_NDK path at line 3 in build_arm_android_64.sh, eg: ANDROID_NDK=/Users/username/path/to/Android-ndk-r14b
3. run ./build_arm_android_64.sh, which will generate the executable file demo.out under build_arm_android_64 folder.
4. run ./test_arm_android_64.sh, which will push demo.out to your android device and execute it.

Notes

1. You can change CMake compilation options in build.sh as needed (i.e. turn off demo/quantools/evaluation/converter/test/benchmark options, turn on openmp/opencl/opengl/vulkan as your backend, set FP16/BF16 low precision mode, and etc). Check MNN document and MNN CMakeList for more information.
2. MNN compilation artifacts under build/mnn_binary_dir
   • libMNN: Backend Shared Library
   • libMNNTrain: Training Framework Shared Library
   • libMNNExpr: Express Training API Shared Library
3. To run demo.out on your linux/macOS machine, first download MNIST dataset from Google Drive, and run ./demo.out /path/to/data/mnist_data. To run demo.out on your android device, adb push mnist_data to your android device under /data/local/tmp before running ./test_arm_android_64.sh.

Dependency

MNN: https://github.com/FedML-AI/MNN.git

pytorch: https://github.com/FedML-AI/pytorch.git