/MegEngine

MegEngine 是一个快速、可拓展、易于使用且支持自动求导的深度学习框架

Primary LanguageC++OtherNOASSERTION

MegEngine

MegEngine Logo

English | 中文

MegEngine is a fast, scalable and easy-to-use deep learning framework, with auto-differentiation.


Installation

NOTE: MegEngine now only supports Linux platform with Python 3.5 or higher. On Windows 10 you could try WSL(Windows Subsystem for Linux) to use Linux within Windows.

Binaries

Commands to install from binaries via pip wheels are as follows:

pip3 install megengine -f https://megengine.org.cn/whl/mge.html

Build from Source

Prerequisites

Most of the dependencies of MegEngine are located in third_party directory, and you do not need to install these by yourself. you can prepare these repositories by executing:

./third_party/prepare.sh
./third_party/install-mkl.sh

But some dependencies should be manually installed:

  • CUDA(>=10.1), cuDNN(>=7.6)are required when building MegEngine with CUDA support (default ON)
  • TensorRT(>=5.1.5) is required when building with TensorRT support (default ON)
  • LLVM/Clang(>=6.0) is required when building with Halide JIT support (default ON)
  • Python(>=3.5), Numpy, SWIG(>=3.0) are required to build Python modules. (default ON)

Build

MegEngine prefers Out-Of-Source flavor, and compile in a mostly-static way. Here are the instructions:

  1. Make a directory for the build.

    mkdir -p build
    cd build
  2. Generate build configurations by CMake.

    For CUDA build:

    cmake .. -DMGE_WITH_TEST=ON

    For CPU only build, use -DMGE_WITH_CUDA=OFF:

    cmake .. -DMGE_WITH_CUDA=OFF -DMGE_WITH_TEST=ON

    For deployment with C++ only, use -DMGE_INFERENCE_ONLY=ON, and turn off test with -DMGE_WITH_TEST=OFF:

    cmake .. -DMGE_INFERENCE_ONLY=ON -DMGE_WITH_TEST=OFF

    Use -DCMAKE_INSTALL_PREFIX=YOUR_PATH to specify the install path.

  3. Start to build.

    make -j$(nproc)
  4. [optional] Install the library if compiled for deployment at step 2.

    make install

Here are some other useful options for the build.

  • MGE_ARCH specifies which arch MegEngine are building for. (default AUTO)
  • MGE_WITH_DISTRIBUTED if multiple machine distributed support is enabled. (default ON)
  • MGE_WITH_PYTHON_MODULE if build python module. (default ON)
  • MGE_BLAS chooses MKL or OpenBLAS as BLAS library for MegEngine. (default MKL)
  • MGE_CUDA_GENCODE supplies the -gencode option for nvcc. (default not supply)
  • MGE_DISABLE_FLOAT16 if disable float16 support. (default OFF)
  • MGE_ENABLE_EXCEPTIONS if enable exception support in C++. (default ON)
  • MGE_ENABLE_LOGGING if enable logging in MegEngine. (default AUTO)

More options can be found by:

cd build
cmake -LAH .. 2>/dev/null| grep -B 1 'MGE_' | less

How to Contribute

We believe we can build an open and friendly community and power humanity with AI.

How to contact us

Resources

License

MegEngine is Licensed under the Apache License, Version 2.0

Copyright (c) 2014-2020 Megvii Inc. All rights reserved.