[
https://issues.apache.org/jira/browse/SOLR-11838?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16344095#comment-16344095
]
Adam Gibson commented on SOLR-11838:
------------------------------------
[~cpoerschke] Yes we would be more than glad to accommodate you here.
That model import module is mainly for keras import. In nd4j (our tensor
library) we are working on supporting both the onnx file format
([http://onnx.ai|http://onnx.ai%28/]) as well as the tensorflow fromat as well.
I think [~jeroens] has the right idea re: dl4j. I can say that it has special
requirements that are pretty atypical of infrastructure typical of running
solr. This is especially true with batch inference.
[~joel.bernstein] SOMs are far from interesting in 2018. Dl4j has Variational
AutoeEncoders which are both the cutting edge right before GANs but also more
stable. Variational AutoEncoders , with their built in reconstruction mechanic
could also be used to track anomalies or rank results right in solr. We use
this workflow mainly for time series data but due to solr being a ranking
engine it makes a ton of sense here.
Re: Time series/LSTMs.The vectors themselves are 3d, but are created from a
time dimension with a specific window length. We typically create this with
what we call a SequenceRecordReader which can take varying file inputs and
infer a time series length based on the input. For solar, a timestamp field
could be used to create the proper 3d vectors pretty easily.
Re: Datavec record reader. We are happy to take pull requests. It is definitely
of interest thanks!
> explore supporting Deeplearning4j NeuralNetwork models
> ------------------------------------------------------
>
> Key: SOLR-11838
> URL: https://issues.apache.org/jira/browse/SOLR-11838
> Project: Solr
> Issue Type: New Feature
> Reporter: Christine Poerschke
> Priority: Major
> Attachments: SOLR-11838.patch, SOLR-11838.patch
>
>
> [~yuyano] wrote in SOLR-11597:
> bq. ... If we think to apply this to more complex neural networks in the
> future, we will need to support layers ...
> [~malcorn_redhat] wrote in SOLR-11597:
> bq. ... In my opinion, if this is a route Solr eventually wants to go, I
> think a better strategy would be to just add a dependency on
> [Deeplearning4j|https://deeplearning4j.org/] ...
> Creating this ticket for the idea to be explored further (if anyone is
> interested in exploring it), complimentary to and independent of the
> SOLR-11597 RankNet related effort.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
