Android Guide

Use FeatherCNN in an Android Studio Project

This is a beginners' example to use FeatherCNN in Android Studio 3.1. We start with a blank project which checked "C++ support". Android Studio will generate a project using Gradle and CMake. For the Android.mk case, please refer to this site. In the following, we illustrate necessary steps to add models and libraries to the project.

Add library to the project

• In the project navigator, expand the file hierarchy and find the libs folder.
• Right click or hold control key and click (macOS) on libs and open this folder in the popup window.
• Copy the compiled library to this folder. If it is compiled with source code, find your library in build-android/install/feather.
• After copy, you should be able to see the library in the navigator.

Edit CMakeLists.txt

• Open CMakeLists.txt
• Edit the lines as illustrated in the figure.
• The target_include_directories tells cmake where to find lib headers.
• The target_link_libraries part links our library with your project.

Edit build.gradle

• Open build.gradle file in the module, NOT that one for the project.
• Add -fopenmp flag for the project in case link errors.
• Add abiFilters to tell Android Studio only compile supported ABIs.

Your project should be able to build with FeatherCNN now.

Add model to the resources

We need to set up a raw folder in the resources.

Just open the File Explorer or Finder to copy your model here.

JNI function calls

In current Android Studio version (3.1), you can see a cpp folder with native-libs.cpp in it. Lets begin with the JNI code.

• In the C++ side, it needs to accept a byte[] from the Java code with the feathermodel data.
• We copy the data from the JNI, and initialize our net.
• Note that this code snippet put the forward_net in a local variable. In real production situations, you should make it persistent through function calls.

The C++ code snippet:

extern "C"
JNIEXPORT void
JNICALL
Java_com_tencent_lan_feather_MainActivity_initFeatherNet(
		JNIEnv *env,
		jobject,
		jbyteArray modelData) {
	jsize modelDataLen = env->GetArrayLength(modelData);
	jbyte *modelDataElem = env->GetByteArrayElements(modelData, NULL);
	char *buffer = (char*) malloc(sizeof(char) * modelDataLen);
	memcpy(buffer, modelDataElem, sizeof(char) * modelDataLen);
	env->ReleaseByteArrayElements(modelData, modelDataElem, 0);
	/* Init feather net */
	int num_threads = 1;
	feather::Net net(num_threads);
	net.InitFromBuffer(buffer);
}

• At the Java side, we get the feathermodel from resource manager and fill that into a single byte array.
• Pass that byte array to the JNI function.
• Don't forget to declare the JNI interface.

The code snippet

InputStream is = getResources().openRawResource(R.raw.mobilenet);
try{
	byte[] modelData = new byte[is.available()];
	is.read(modelData);
	initFeatherNet(modelData);
} catch (IOException e){
	//Handle file exception here.
}

Now the net is properly initialized. You can perform inference computation with the Forward method through JNI.