[
https://issues.apache.org/jira/browse/IGNITE-10202?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16684988#comment-16684988
]
Anton Dmitriev edited comment on IGNITE-10202 at 11/13/18 10:27 AM:
--------------------------------------------------------------------
The following repository has been prepared:
[https://github.com/dmitrievanthony/tensorflow-inference-on-apache-ignite].
It contains two examples of using TensorFlow model for inference from Java. One
is an inference on classical MNIST modes, another is an inference based on face
detection model.
Using HDF5 (.h5) format doesn't look promising so far because it's not
integrated with TensorFlow Estimators well enough.
was (Author: dmitrievanthony):
The following repository has been prepared:
[https://github.com/dmitrievanthony/tensorflow-inference-on-apache-ignite].
It contains two examples of using TensorFlow model for inference from Java. One
is an inference on classical MNIST modes, another is an inference based on face
detection model.
Using HDF5 format doesn't look promising so far because it's not integrated
with TensorFlow Estimators well enough.
> ML: Create a POC of TensorFlow model inference in Java
> ------------------------------------------------------
>
> Key: IGNITE-10202
> URL: https://issues.apache.org/jira/browse/IGNITE-10202
> Project: Ignite
> Issue Type: Sub-task
> Components: ml
> Affects Versions: 2.8
> Reporter: Anton Dmitriev
> Assignee: Anton Dmitriev
> Priority: Major
> Fix For: 2.8
>
>
> TensorFlow models are represented internally as computation graphs that are
> saved as a Protobuf objects with some kind of metadata. TensorFlow provides a
> good API for Python that allows to work with them, but Java API is very poor.
> TensorFlow allows to import computation graph in Java API and manipulate it
> using low-level API and it looks like a one way to make model inference in
> Java. Another way is based on the fact that the most popular way to build a
> TensorFlow model is using of the Keras API. It allows to save model in _.h5_
> format that can be imported into Java using Deeplearning4J.
> So, the goal of this task is to *prepare a POC that demonstrates*:
> * How to import TensorFlow model into Java and make inference using
> TensorFlow Java API.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
