README.md

September 26, 2018 · View on GitHub

Future development of this binding has been moved to nexus-torch (https://github.com/ctongfei/nexus/tree/master/torch), serving as a computation backend of nexus, a typesafe deep learning engine.

JTorch: JVM bindings for Torch (PyTorch C/C++ core)

Corresponding PyTorch version: 0.4.0

🚧 Ongoing project 🚧 Status: Not ready for use 🚧

	Windows	Linux	MacOS
CPU			✓
CUDA			✓

Installation

Step 1: Install PyTorch from source.

Locate the headers (*.h) and shared libraries (*.so / *.dylib / *.dll)

Step 2: Build SWIG bindings and publish to local Ivy repository.

./build-{mac/linux}.sh

You probably need to modify something in the script to make it work.

[TODO] Will be refactored into a Makefile.

Step 3: Use!

    <dependency>
      <groupId>me.tongfei</groupId>
      <artifactId>jtorch-java</artifactId>
      <version>0.1-TH0.4-SNAPSHOT</version>
    </dependency>

When starting a JVM, add -Djava.library.path=/usr/local/lib (or where your libATen.dylib is), or set $LD_LIBRARY_PATH to include that directory, so that JVM can properly load native the binary.

Package mappings

Torch has several C packages, to which each is mapped to the following Java package:

Torch package	Mapped Java package
`TH`	`jtorch`
`THNN`	`jtorch.nn`
`THS`	`jtorch.sparse`
`THC`	`jtorch.cuda`
`THCUNN`	`jtorch.cuda.nn`
`THCS`	`jtorch.cuda.sparse`

Type mappings

Torch has the following Storage/Tensor types, which are mapped to the Java corresponding types on the right.

Torch type	Element type (C)	Bits	Mapped element type (Java)	Java type
`THHalfTensor`	`THHalf`	16	`float`	`HalfTensor`
`THFloatTensor`	`float`	32	`float`	`FloatTensor`
`THDoubleTensor`	`double`	64	`double`	`DoubleTensor`
`THByteTensor`	`uint8_t`	8	`short`	`UByteTensor`
`THCharTensor`	`int8_t`	8	`byte`	`ByteTensor`
`THShortTensor`	`int16_t`	16	`short`	`ShortTensor`
`THIntTensor`	`int32_t`	32	`int`	`IntTensor`
`THLongTensor`	`int64_t`	64	`long`	`LongTensor`