DJL MXNet Android Template¶

Introduction¶

This template includes 1. how you build the mxnet native library(libmxnet.so) for android 2. how you use MXNet native library(libmxnet.so) in Android Studio with DJL to run a simple image classification. DJL team also prepare an unofficial prebuilt libmxnet.so on 4 common Android ABI. You can take advantage of those and skip the step 1.

Android ABI
armeabi-v7a
arm64-v8a
x86
x86_64

1. Build MXNet from source on Android platform¶

MXNet Android uses OpenBLAS to do matrix computation. Before we build MXNet from source, we need to get libopenblas.so ready on common Android ABI which are armeabi-v7a, arm64-v8a, x86 and x86_64 respectively.

Build OpenBlas from source¶

Here is the minimal script to build openblas assuming you have a base ubuntu 20.02 image.

Prerequisite¶

apt update && apt install -y \
    build-essential \
    ninja-build \
    cmake \
    ccache \
    git \
    curl \
    unzip

# We use NDK r19 version here and use /usr/local as home dir
# You can switch to newer version and other dirs
cd /usr/local
curl -o android-ndk-r19-linux-x86_64.zip -L https://dl.google.com/android/repository/android-ndk-r19-linux-x86_64.zip
unzip android-ndk-r19-linux-x86_64.zip
rm android-ndk-r19-linux-x86_64.zip
export CMAKE_TOOLCHAIN_FILE=/usr/local/android-ndk-r19/build/cmake/android.toolchain.cmake
export TOOLCHAIN=/usr/local/android-ndk-r19/toolchains/llvm/prebuilt/linux-x86_64

Build OpenBLAS on different arch¶

# latest version
git clone --recursive -b v0.3.12 https://github.com/xianyi/OpenBLAS.git
mkdir /usr/local/openblas-android
cd /usr/local/OpenBLAS

# Build armv7
make \
    TARGET=ARMV7 \
    ONLY_CBLAS=1 \
    CC="$TOOLCHAIN"/bin/armv7a-linux-androideabi21-clang \
    AR="$TOOLCHAIN"/bin/arm-linux-androideabi-ar \
    HOSTCC=gcc \
    ARM_SOFTFP_ABI=1 \
    -j4
make PREFIX=/usr/local/openblas-android NO_SHARED=1 install

# Build armv8
make \
    TARGET=CORTEXA57 \
    ONLY_CBLAS=1 \
    CC=$TOOLCHAIN/bin/aarch64-linux-android21-clang \
    AR=$TOOLCHAIN/bin/aarch64-linux-android-ar \
    HOSTCC=gcc \
    -j4
make PREFIX=/usr/local/openblas-android NO_SHARED=1 install

# Build x86
make \
    TARGET=ATOM \
    ONLY_CBLAS=1 \
    CC="$TOOLCHAIN"/bin/i686-linux-android21-clang \
    AR="$TOOLCHAIN"/bin/i686-linux-android-ar \
    HOSTCC=gcc \
    ARM_SOFTFP_ABI=1 \
    -j4
make PREFIX=/usr/local/openblas-android NO_SHARED=1 install

# Build x86_64
make \
    TARGET=ATOM BINARY=64\
    ONLY_CBLAS=1 \
    CC="$TOOLCHAIN"/bin/x86_64-linux-android21-clang \
    AR="$TOOLCHAIN"/bin/x86_64-linux-android-ar \
    HOSTCC=gcc \
    ARM_SOFTFP_ABI=1 \
    -j4
make PREFIX=/usr/local/openblas-android NO_SHARED=1 install

# This will tell mxnet where to find the openblas
export OpenBLAS_HOME=/usr/local/openblas-android

Build MXNet from source¶

git clone --recursive -b 1.8.0 https://github.com/apache/incubator-mxnet.git
mkdir build
cd build

# Build armv7
cmake \
        -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} \
        -DANDROID_ABI="armeabi-v7a" \
        -DANDROID_STL="c++_static" \
        -DANDROID=ON \
        -DUSE_CUDA=OFF \
        -DUSE_SSE=OFF \
        -DUSE_LAPACK=OFF \
        -DUSE_OPENCV=OFF \
        -DUSE_OPENMP=OFF \
        -DUSE_SIGNAL_HANDLER=ON \
        -DUSE_MKL_IF_AVAILABLE=OFF \
        -G Ninja ..
ninja

# Build armv8
cmake \
        -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} \
        -DANDROID_ABI="arm64-v8a" \
        -DANDROID_STL="c++_static" \
        -DANDROID=ON \
        -DUSE_CUDA=OFF \
        -DUSE_SSE=OFF \
        -DUSE_LAPACK=OFF \
        -DUSE_OPENCV=OFF \
        -DUSE_OPENMP=OFF \
        -DUSE_SIGNAL_HANDLER=ON \
        -DUSE_MKL_IF_AVAILABLE=OFF \
        -G Ninja ..
ninja

# Build x86
cmake \
        -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} \
        -DANDROID_ABI="x86" \
        -DANDROID_STL="c++_static" \
        -DANDROID=ON \
        -DUSE_CUDA=OFF \
        -DUSE_SSE=OFF \
        -DUSE_LAPACK=OFF \
        -DUSE_OPENCV=OFF \
        -DUSE_OPENMP=OFF \
        -DUSE_SIGNAL_HANDLER=ON \
        -DUSE_MKL_IF_AVAILABLE=OFF \
        -G Ninja ..
ninja

# Build x86-64
cmake \
        -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TOOLCHAIN_FILE} \
        -DANDROID_ABI="x86_64" \
        -DANDROID_STL="c++_static" \
        -DANDROID=ON \
        -DUSE_CUDA=OFF \
        -DUSE_SSE=OFF \
        -DUSE_LAPACK=OFF \
        -DUSE_OPENCV=OFF \
        -DUSE_OPENMP=OFF \
        -DUSE_SIGNAL_HANDLER=ON \
        -DUSE_MKL_IF_AVAILABLE=OFF \
        -G Ninja ..
ninja

If you don't run into any error, you can see libmxnet.so in the build folder

2. Android Studio Setup¶

Depending on Android you would like to test, you just need to put the corresponding libmxnet.so under the src/main/jniLibs/{ABI}. For example, if you want to run on armeabi-v7a, put the right libmxnet.so under src/main/jniLibs/armeabi-v7a.

Then you're all set. You can run this project and should be able to see the classification result right in the middle.

Here are extra setting I did when I created the project.

In build.gradle, I included 4 packages ai.djl:api, ai.djl.android:core, ai.djl.mxnet:mxnet-engine and ai.djl.mxnet:mxnet-model-zoo to use DJL API. Because DJL MXNet interact with MXNet engine via JNA, one more dependency called net.java.dev.jna:jna is required as well.
DJL API depends on JNA by default. For Android project, we should use net.java.dev.jna:jna:5.3.0@aar instead. We need to exclude duplicate JNA dependencies by using exclude group: "net.java.dev.jna", module: "jna"
Android system only allows certain location where you can download files from internet. DJL by default will choose ~/.djl.ai as directory for keeping models. To make sure we download the model into the right place, we expose a System Property DJL_CACHE_DIR.
The project involves downloading the model from internet, so I add <uses-permission android:name="android.permission.INTERNET" /> and move model loading in AsyncTask.

In terms of DJL API usage, you can find more details on Official Doc.