Add PlayTorch to Existing App
In this tutorial, you will learn how to add the PlayTorch core package to an existing React Native project.
If you have an existing React Native project and you want to add ML capabilities, you can add the react-native-pytorch-core
package. This package includes all code needed to run ML inference, the Canvas
, Camera
, ImageUtil
, and AudioUtil
.
Installation
yarn add react-native-pytorch-core
caution
If you are on React Native 0.68+, you will have to use the nightly
version of the core package. A fix has been added in commit fix RN0.68 can't build with react-native-pytorch-core, which is currently only available in the nightly
version.
yarn add react-native-pytorch-core@nightly
On iOS you are done, but Android requires the following additional steps for the react-native-pytorch-core
package to work.
Additional Assets for Metro
If the PyTorch Mobile models are part of the React Native bundle, the Metro configuration needs to be changed to resolve the ptl
files.
note
This is only required if models are loaded from the bundle using require('./path/to/model.ptl')
. It is not required if models are loaded from the mobile device's local file system or via a URL.
// get defaults assetExts array
const defaultAssetExts = require('metro-config/src/defaults/defaults')
.assetExts;
module.exports = {
// ...
resolver: {
assetExts: [...defaultAssetExts, 'ptl'],
},
// ...
};
Additional steps on Android
For the react-native-pytorch-core
React Native package to work on Android, it requires three changes to the gradle.properties
and the two build.gradle
files to increase JVM memory, add Sonatype repository, and packaging options with pick first rule.
Increase JVM Memory
Increase the memory for the JVM to avoid OutOfMemory exceptions during the packaging process.
org.gradle.jvmargs=-Xmx4g
Without the increased memory, the packaging process might fail with the following error:
* What went wrong:
Execution failed for task ':app:packageDebug'.
> A failure occurred while executing com.android.build.gradle.tasks.PackageAndroidArtifact$IncrementalSplitterRunnable
> java.lang.OutOfMemoryError (no error message)
Sonatype Repository
The PyTorch Mobile for Android dependencies are in the Sonatype repository. Add the repository url to the allprojects > repositories
.
allprojects {
repositories {
// ...
maven {
url("https://oss.sonatype.org/content/repositories/snapshots")
}
// ...
}
}
Update app build.gradle
Both React Native and PyTorch Mobile for Android use fbjni
. For example, the versions for PlayTorch that are used for development are:
- React Native
0.64.3
uses fbjni0.0.2
- PyTorch Mobile
1.12.2
uses fbjni0.2.2
.
So far, fbjni
is forward compatible, which means it is ok to pick the latest version shipped with either of the two dependencies. At this point, it is fbjni 0.2.2
. For Gradle to pick the right version, the android/app/build.gradle
needs to have a few adjustments:
- Add
pickFirst
rule topackagingOptions
. This rule will pick the first shared object (dynamic) library. It will give higher priority to shared object libraries that are coming with direct app dependencies, which is why 2. is important. - Set up an extra directory for fbjni where the fbjni version
0.2.2
from the dependency added in 3. will be extracted. Also add the relevant task to thebuild.gradle
file (seetask extraJNILibs
andtasks.whenTaskAdded
after thedependencies
definition) - Add fbjni
0.2.2
as direct app dependency.
See the build.gradle
in the react-native-pytorch-core
example app for a possible configuration.
info
The following error will show if pickFirst
is not set:
* What went wrong:
Execution failed for task ':app:mergeDebugNativeLibs'.
> A failure occurred while executing com.android.build.gradle.internal.tasks.Workers$ActionFacade
> More than one file was found with OS independent path 'lib/x86/libfbjni.so'
android {
// ...
/**
* Without the packaging options, it will result in the following build error:
*
* * What went wrong:
* Execution failed for task ':app:mergeDebugNativeLibs'.
* > A failure occurred while executing com.android.build.gradle.internal.tasks.Workers$ActionFacade
* > More than one file was found with OS independent path 'lib/x86/libfbjni.so'
*/
packagingOptions {
pickFirst '**/*.so'
}
sourceSets {
main {
jniLibs.srcDirs += ["$buildDir/extra-jniLibs/jni"]
}
}
configurations {
extraJNILibs
}
// ...
}
dependencies {
// ...
// Used to control the version of libfbjni.so packaged into the APK
extraJNILibs("com.facebook.fbjni:fbjni:0.2.2")
// ...
}
// ...
// Extract JNI shared libraries as project libraries. This assumes the target directory, $buildDir/extra-jniLibs, is added to the jniLibs.srcDirs configuration.
task extraJNILibs {
doLast {
configurations.extraJNILibs.files.each {
def file = it.absoluteFile
copy {
from zipTree(file)
into "$buildDir/extra-jniLibs" // temp location instead of "src/main/jniLibs"
include "jni/**/*"
}
}
}
}
tasks.whenTaskAdded { task ->
if (task.name == 'mergeDebugJniLibFolders' || task.name == 'mergeReleaseJniLibFolders') {
task.dependsOn(extraJNILibs)
}
}
// ...
Configure JavaScript Interface for React Native app
Last, add the PyTorchCoreJSIModulePackage
to the ReactNativeHost
in the app's MainApplication.java
.
info
JavaScript Interface (JSI)
A lightweight and VM-independent layer that helps in communication between JavaScript and native platforms easier and faster. It supports Webkit JSC, Custom JSC, and Hermes.
JSI Documentation
The jsi.h in the Hermes GitHub repository and has a really well-documented C++ header, which is a great place to learn more about JSI.
PlayTorch uses the JavaScript Interface (JSI) to expose PyTorch Mobile C++ functions
import com.facebook.react.bridge.JSIModulePackage;
import org.pytorch.rn.core.jsi.PyTorchCoreJSIModulePackage;
// ...
new ReactNativeHostWrapper(
this,
new ReactNativeHost(this) {
// ...
@Override
protected JSIModulePackage getJSIModulePackage() {
return new PyTorchCoreJSIModulePackage();
}
});
Great! You are done if you don't use other React Native packages that rely on JSI, otherwise continue!
Configure multiple React Native packages using JSI
If you use other React Native packages with JSI packages, you will have to create a JSIModulePackage
that combines all of them. For example, if you use React Reanimated and PlayTorch Core create a CustomJSIModulePackage
and return this in the ReactNativeHost
implementation.
package <your-package-path>;
import androidx.annotation.Keep;
import com.facebook.proguard.annotations.DoNotStrip;
import com.facebook.react.bridge.JSIModulePackage;
import com.facebook.react.bridge.JSIModuleSpec;
import com.facebook.react.bridge.JavaScriptContextHolder;
import com.facebook.react.bridge.ReactApplicationContext;
import com.facebook.soloader.SoLoader;
import com.swmansion.reanimated.ReanimatedJSIModulePackage;
import java.util.List;
import org.pytorch.rn.core.jsi.PyTorchCoreJSIModulePackage;
/**
* This is a custom JSIModulePackage that combines ReanimatedJSIModulePackage and
* PyTorchCoreJSIModulePackage. It is needed because the ReactNativeHostWrapper only supports
* returning a single JSIModulePackage.
*/
@DoNotStrip
@Keep
public class CustomJSIModulePackage implements JSIModulePackage {
@DoNotStrip
@Keep
@Override
public List<JSIModuleSpec> getJSIModules(
ReactApplicationContext reactApplicationContext, JavaScriptContextHolder jsContext) {
ReanimatedJSIModulePackage reaJSIModulePackage = new ReanimatedJSIModulePackage();
PyTorchCoreJSIModulePackage ptlJSIModulePackage = new PyTorchCoreJSIModulePackage();
List<JSIModuleSpec> retList =
reaJSIModulePackage.getJSIModules(reactApplicationContext, jsContext);
ptlJSIModulePackage.getJSIModules(reactApplicationContext, jsContext);
return retList;
}
}
// ...
import com.facebook.react.bridge.JSIModulePackage;
new ReactNativeHostWrapper(
this,
new ReactNativeHost(this) {
// ...
@Override
protected JSIModulePackage getJSIModulePackage() {
// The CustomJSIModulePackage combines JSIModulePackage from different React Native
// packages (e.g., Reanimated2 and PlayTorch).
return new CustomJSIModulePackage();
}
});
That should be all necessary Gradle build configuration changes!