[
https://issues.apache.org/jira/browse/SOLR-11838?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16321580#comment-16321580
]
Adam Gibson commented on SOLR-11838:
------------------------------------
Hi Gus,
There are 2 sources of off heap memory to consider.
We cover that here:
http://deeplearning4j.org/memory
For the GPU we have essentially our own GC. The thing to think about here if
you want to have either cpu or gpu, is both are optional.
I'm more than glad to answer questions on that if anyone has any concerns.
Beyond that, Nd4j itself is very similar to slf4j. You pick the "chip" you
want as a jar file.
So you could in theory have 2 class paths, one for cpu and one for gpu, picking
one as the default.
Pretrain models can be either a computation graph or a multi layer network. We
have a ModelGuesser that helps mitigate the various types.
We are introducing another type as well soonish that can directly import
tensorflow and onnx as well (this will be a more flexible api similar to
pytorch)
which will also work. We will be releasing that within the next few weeks.
Depending on the timelines for the release, we're happy to coordinate with
folks interested
in various pretrained models.
This is in top of our existing keras support.
For the untrained network/various hyper parameters, might I suggest allowing
folks to upload a config of their choice? You can try to offer various kinds of
sample architectures but we've found that the best way to handle this in
practice is by just allowing folks to upload their own architectures.
For the datasetiterator: That is mainly used for minibatch training. You can
also create datasets on the fly as well.
For inference purposes, dl4j makes no assumptions. You could technically just
call network.output an on INDArray directly)
The solr project might be also interested in our alpha sparse support if you
need to convert a document vector directly for inference purposes.
> explore supporting Deeplearning4j NeuralNetwork models in contrib/ltr
> ---------------------------------------------------------------------
>
> Key: SOLR-11838
> URL: https://issues.apache.org/jira/browse/SOLR-11838
> Project: Solr
> Issue Type: New Feature
> Components: contrib - LTR
> Reporter: Christine Poerschke
> Attachments: 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
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
