onnxruntime/java
Baiju Meswani 10ba1e270c
Minimal Build for On-Device Training (#16326)
🛠️ __Changes in this pull request:__

This pull request introduces two significant changes to the project:

- Changing on device training checkpoint format: The current
implementation stores the on device training checkpoint as a sequence of
tensors in multiple files inside a checkpoint folder, which can be
inefficient in terms of storage and performance. In this PR, I have
modified the checkpoint format to utilize the flatbuffer table to save
the checkpoint to a single file, providing a more compact and efficient
representation. The changes around this are twofold:
- Add the checkpoint flatbuffer schema that will generate the necessary
checkpoint source files.
- Update the checkpoint saving and loading functionality to use the new
format.

- Adding support for onnxruntime minimal build: To support scenarios
where binary size is a constraint, I made changes to ensure that the
training build can work well with the minimal build.

🔍 __Open Issues:__
- In order to extract the optimizer type, the existing implementation
re-loaded the onnx optimizer model and parsed it. This is no longer
possible, since the model format can either be onnx or ort. One idea is
to do the same for ort format optimizer model. This needs some
investigation.
- Changes to the offline tooling to generate ort format training
artifacts.
- End-to-end training example showcasing the use of the minimal training
build.
- Add support for export model for inferencing in a minimal build.
2023-06-22 12:27:23 -07:00
..
gradle/wrapper Update Gradle version (#14862) 2023-03-08 12:22:06 -08:00
src Minimal Build for On-Device Training (#16326) 2023-06-22 12:27:23 -07:00
testdata [java] Sparse tensor support (#10653) 2022-11-22 10:29:24 -08:00
build-android.gradle Creating Nuget and Android packages for Training (#15712) 2023-05-01 12:59:56 -07:00
build.gradle Fixing CoreML in Java (#16231) 2023-06-07 12:24:57 -07:00
gradlew Update Gradle version (#14862) 2023-03-08 12:22:06 -08:00
gradlew.bat Update Gradle version (#14862) 2023-03-08 12:22:06 -08:00
README.md Update Gradle version (#14862) 2023-03-08 12:22:06 -08:00
settings-android.gradle
settings.gradle

ONNX Runtime Java API

This directory contains the Java language binding for the ONNX runtime. Java Native Interface (JNI) is used to allow for seamless calls to ONNX runtime from Java.

Usage

This document pertains to developing, building, running, and testing the API itself in your local environment. For general purpose usage of the publicly distributed API, please see the general Java API documentation.

Building

Use the main project's build instructions with the --build_java option.

Requirements

JDK version 8 or later is required.

The Gradle build system is used here to manage the Java project's dependency management, compilation, testing, and assembly. In particular, the Gradle wrapper at java/gradlew[.bat] is used, locking the Gradle version to the one specified in the java/gradle/wrapper/gradle-wrapper.properties configuration. Using the Gradle wrapper removes the need to have the right version of Gradle installed on the system.

Build Output

The build will generate output in $REPO_ROOT/build/$OS/$CONFIGURATION/java/build:

  • docs/javadoc/ - HTML javadoc
  • reports/ - detailed test results and other reports
  • libs/onnxruntime-VERSION.jar - JAR with compiled classes, platform-specific JNI shared library, and platform-specific onnxruntime shared library.

Build System Overview

The main CMake build system delegates building and testing to Gradle. This allows the CMake system to ensure all of the C/C++ compilation is achieved prior to the Java build. The Java build depends on C/C++ onnxruntime shared library and a C JNI shared library (source located in the src/main/native directory). The JNI shared library is the glue that allows for Java to call functions in onnxruntime shared library. Given the fact that CMake injects native dependencies during CMake builds, some gradle tasks (primarily, build, test, and check) may fail.

When running the build script, CMake will compile the onnxruntime target and the JNI glue onnxruntime4j_jni target and expose the resulting libraries in a place where Gradle can ingest them. Upon successful compilation of those targets, a special Gradle task to build will be executed. The results will be placed in the output directory stated above.

Advanced Loading

The default behavior is to load the shared libraries using classpath resources. If your use case requires custom loading of the shared libraries, please consult the javadoc in the package-info.java or OnnxRuntime.java files.

Development

Code Formatting

Spotless is used to keep the code properly formatted. Gradle's spotlessCheck task will show any misformatted code. Gradle's spotlessApply task will try to fix the formatting. Misformatted code will raise failures when checks are ran during test run.

JNI Headers

When adding or updating native methods in the Java files, it may be necessary to examine the relevant JNI headers in build/headers/ai_onnxruntime*.h. These files can be manually generated using Gradle's compileJava task which will compile the Java and update the header files accordingly. Then the corresponding C files in ./src/main/native/ai_onnxruntime*.c may be updated and the build can be ran.

Dependencies

The Java API does not have any runtime or compile dependencies currently.