[IGNITE-8335] TensorFlow integration

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Done
• Fix Version/s: 2.7
• Component/s: ml
• Assignee: Anton Dmitriev
• Reporter: Yury Babak
• Created: 20/Apr/18 08:47
• Updated: 10/Oct/18 17:16
• Resolved: 10/Oct/18 17:16

Description

The goal of TensorFlow on Apache Ignite is to allow users to train and inference neural network models on a data stored in Apache Ignite distributed database utilizing all TensorFlow functionality and Apache Ignite data collocation abilities.

There are 8 questions we need to answer to build TensorFlow on Apache Ignite:

• How to build and maintain TensorFlow cluster on top of Apache Ignite infrastructure utilizing Apache Ignite data collocation abilities?
• How to pass data from Apache Ignite storage into TensorFlow?
• How to organize load balancing and optimize Apache Ignite data distribution to improve training performance?
• How to integrate TensorFlow checkpoints mechanism into Apache Ignite ecosystem?
• How to recover after cluster node failures?
• How to deploy TensorFlow on Apache Ignite in local/cluster/cloud environment?
• How to serve model built in TensorFlow on Apache Ignite?
• How to profile training using TensorBoard or similar tools?

Issue Links

• is part of: IGNITE-8670 Umbrella: TensorFlow integration
• relates to: IGNITE-7437 Partition based dataset implementation
• relates to: IGNITE-7782 Thin Client lib: Python
• relates to: IGNITE-4600 Apache Ignite Python Library
• links to: Issue in TensorFlow repository
• links to: TensorFlow on Apache Ignite
• links to: Topic on TensorFlow developers mail list
• links to: Topic on TensorFlow discussion mail list