/dl-on-flink

Primary LanguageJavaApache License 2.0Apache-2.0

Deep Learning on Flink

Deep Learning on Flink aims to integrate Flink and deep learning frameworks (e.g. TensorFlow, PyTorch, etc.) to enable distributed deep learning training and inference on a Flink cluster.

It runs the deep learning tasks inside a Flink operator so that Flink can help establish a distributed environment, manage the resource, read/write the data with the rich connectors in Flink and handle the failures.

Currently, Deep Learning on Flink supports TensorFlow.

Supported Operating System

Deep Learning on Flink is tested and supported on the following 64-bit systems:

  • Ubuntu 18.04
  • macOS 10.15

Support Framework Version

  • TensorFlow: 1.15.x & 2.4.x
  • PyTorch: 1.11.x
  • Flink: 1.14.x

Getting Started

Deep learning on Flink currently works with Tensorflow and PyTorch. You can see the following pages for the usage and examples.

Build From Source

Requirements

  • python: 3.7
  • cmake >= 3.6
  • java 1.8
  • maven >=3.3.0

Deep Learning on Flink requires Java and Python works together. Thus, we need to build for both Java and Python.

Initializing Submodules before Building Deep Learning on Flink from Source

Please use the following command to initialize submodules before building from source.

git submodule update --init --recursive

Build Java

mvn -DskipTests clean install

After finish, you can find the target distribution in the dl-on-flink-dist/target folder.

Build Python

Install from Source

You can run the following commands to install the Python packages from source

# Install dl-on-flink-framework first
pip install dl-on-flink-framework/python

# Note that you should only install one of the following as they require
# different versions of Tensorflow 
# For tensorflow 1.15.x
pip install dl-on-flink-tensorflow/python
# For tensorflow 2.4.x
pip install dl-on-flink-tensorflow-2.x/python

Build wheels

We provide a script to build wheels for Python packages, you can run the following command.

bash tools/build_wheel.sh

After finish, you can find the wheels at tools/dist. Then you can install the python package with the wheels.

pip install tools/dist/<wheel>

For More Information

Design document

License

Apache License 2.0