Reply
New Contributor
Posts: 4
Registered: ‎05-08-2015
Accepted Solution

Speed layer in Oryx2

Hi,

 

Is the speed layer following an incremental learning approach in Oryx2? (during Model Update).

 

Thanks

Jayani

Cloudera Employee
Posts: 447
Registered: ‎08-11-2014

Re: Speed layer in Oryx2

The speed layer does indeed produce incremental updates. Like the serving layer, it loads the most recent model in memory and then computes how the model might change (approximately, rapidly) in response to new data, and internalizes and publishes those updates. The serving layers then hear the models but also the updates on the queue and update accordingly.

New Contributor
Posts: 4
Registered: ‎05-08-2015

Re: Speed layer in Oryx2

[ Edited ]

Thanks Sean for the prompt response :)

 

So, can we use or customize the same speed layer approach for mini-batch learning as well?

 

Also, does Oryx have any future plans to support built-in pre-processing methods for text analysis such as Tokenization and TF-IDF vector creation?

 

Thank you

Jayani

Highlighted
Cloudera Employee
Posts: 447
Registered: ‎08-11-2014

Re: Speed layer in Oryx2

Yes, well, I'd say that the batch layer can do "mini batch" if you simply use a low interval time. It's not a special case, really. I think this project isn't going to add its own data prep pipeline, no, but the idea is that you can use any Java or Spark-based libraries you like as part of your app. There's no need to have a different special set of support in this project.

New Contributor
Posts: 4
Registered: ‎05-08-2015

Re: Speed layer in Oryx2

Okay. Thanks Sean for the information.

Explorer
Posts: 74
Registered: ‎07-18-2014

Re: Speed layer in Oryx2

Sean,

Can you provide more detailed information about how the approximation is computed in Oryx 2.0 ?

Is it the same fold-in approach as Oryx 1.0 ? Can you point to the code base as reference ?

 

Thanks.

Jason

Cloudera Employee
Posts: 447
Registered: ‎08-11-2014

Re: Speed layer in Oryx2

ALS: yes, fold-in just as before k-means: assign point to a cluster and update its centroid (but don't reassign any other points) RDF: assign point to leaf and update leaf's prediction (but don't change the rest of the tree)

Announcements