PyTorch Java Demo

This repository is a demonstration of how to use PyTorch from Java.

Setup

Download and unpack libtorch nightly (or 1.4 or greater). From the pytorch.org homepage under "Quick Start Locally", make sure "LibTorch" is the selected package. Linux is supported as of version 1.4. Mac is supported as of version 1.5. Linux and Mac are both supported in nightly builds. Windows is not supported at this time.

Run export LIBTORCH_HOME=/path/to/libtorch. The build.gradle file will use this to set java.library.path when running the application. If you are using PyTorch in your own environment, LIBTORCH_HOME is not necessary. Instead, you will need to set java.library.path to /path/to/libtorch/lib.

If using a nightly build, run export USE_LIBTORCH_NIGHTLY=1.

Run ./gradlew run to build and run the demo application. It will load demo-model.pt and run it on some simple data. This notebook was used to generate the model.

More information about the Java API and TorchScript:

• TorchScript tutorial
• TorchScript reference
• Loading TorchScript in C++
• PyTorch-style JavaDoc
• Standard JavaDoc
• PyTorch Android tutorial

Example Setup for macOS with v 1.7.1

cd ~/code/java
wget https://download.pytorch.org/libtorch/cpu/libtorch-macos-1.7.1.zip
unzip libtorch-macos-1.7.1.zip
export LIBTORCH_HOME="~/code/java/libtorch"
export USE_LIBTORCH_NIGHTLY=1
export JAVA_HOME=$(/usr/libexec/java_home)
./gradlew run