systemml-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Niketan Pansare" <npan...@us.ibm.com>
Subject Re: [PROPOSAL] R4ML Integration with SystemML
Date Fri, 22 Sep 2017 22:10:55 GMT

>> a) Does it mean you are proposing spliting R4ML into two R-wrapper and
R4ML?

I was only suggesting how you ought to stage the PRs into SystemML once the
vote passes :)

>> So I was thinking is it absolutely must have to sync between api?

Soft-yes, we should try our best to do so.

Thanks,

Niketan Pansare
IBM Almaden Research Center
E-mail: npansar At us.ibm.com
http://researcher.watson.ibm.com/researcher/view.php?person=us-npansar



From:	alok singh <singh_alok@hotmail.com>
To:	"dev@systemml.apache.org" <dev@systemml.apache.org>
Date:	09/22/2017 02:40 PM
Subject:	Re: [PROPOSAL] R4ML Integration with SystemML




see comments Alok:



From: Niketan Pansare <npansar@us.ibm.com>
Sent: Friday, September 22, 2017 2:11 PM
To: dev@systemml.apache.org
Subject: Re: [PROPOSAL] R4ML Integration with SystemML

>>> As pointed out earlier, R4ML is not just R interface so it is based on
the earlier product of IBM on R and it has many product feature.
Also note that the pure ML Ctx and the cmd options for dml is not ideally
allow all the things user want to do in his ML code.
The solution could be to create wrapper to make user happy  . but we have
created those wrapper but those are in R and from user point for view it
feels that are just writing the R code
If the ultimate goal is to have just MLCtx based R interface than I think
it undermines and R4ML value proposition.
(We can definitely just expose MLCtx api. However calling Logistic
Regression example just for the purpose of MLCtx won't be best) R4ML.mlogit
has better apis

May be we are not on same page.
(a) MLContext is not the only API, but an important one that needs to be
supported.
(b) Like R4ML, our mllearn wrappers aim to simplify the usage for the
Python users. These wrappers were designed so that if someone wrote a
python script that uses scikit-learn or mllib. Then, a simple change from
`from sklearn import LogisticRegression`  to `from systemml.mllearn import
LogisticRegression` should in principle allow SystemML to be incorporated
in their workflow.

Alok:
a) Does it mean you are proposing spliting R4ML into two R-wrapper and
R4ML? I think that could be idea one can potentially look into it. I second
it. That way one can have pure R wrapper and like mllearn kind of R4ML

b) Currently we can sure expose the MLContext from R as public api but to
use all the code involves many convulations to make life easier for R user.
For example see code func *execute* *output* *getDF* in
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_aloknsingh_r4ml_blob_0d79b3c7975be55989466869fe99ccfd47dd6dc3_R4ML_R_sysml.bridge.R&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=SivKuY8Zn0LQAmM2UmppEwy4L-lROLYUzT9iYnS4Njg&s=VZv9IEtLnaXzZ3mp1bICD4zRv3SL2VO7b68H0wHTCis&e=


>> 1) I think it will require a lot of work for scala and python api to be
in sync with r4ml api.
Also I feel that if the goal is too have just python, scala than we have to
do the coding at R4ML. but I think goals was to merge this project.

I guess the goal is to make SystemML better and more user-friendly. To do
that, we have to try our best to keep our APIs across language consistent.
I understand it might require lot of work for Scala and Python APIs to be
in sync with R4ML API,  but it has to be done.


Since R4ML was designed in isolation with the SystemML project, I am
recommending to do a gradual merge of (1) the additional features and (2)
features that diverge from SystemML APIs so as to be R friendly; thus,
allowing the SystemML community  to comment on them before merging. This
also allows the R4ML features that match one-to-one with the Python and
Scala APIs to be merged quickly and not be in the PR until we agree to
every (1) and (2) features :)

Alok: See the previous comments I like we should explore the idea of
splitting the way you splitted mllearn. Still more discussion needed as I
see it.  At this stage those changes will require complete change at R4ML
to have those.

Another way to think would be that R4ML can be independent package, which
eventually be pushed to CRAN.
note that in the spark dev repo. Spark core is there and SparkR is there as
seperate dir and python is there as seperate dir

Initially, SparkScala, SparkR and pyspark tried to be in sync but I think
now many features are been added which is not causing sync between sparkR
and pyspark and similar between SparkScala and SpakR and  PySpark.

So I was thinking is it absolutely must have to sync between api? Since all
these will cater to different user.

These are ideas.




Thanks,

Niketan Pansare
IBM Almaden Research Center
E-mail: npansar At us.ibm.com
http://researcher.watson.ibm.com/researcher/view.php?person=us-npansar

 http://researcher.watson.ibm.com/researcher/photos/3531.jpg

Niketan Pansare - IBM
researcher.watson.ibm.com
Niketan Pansare is a Senior Software Engineer at IBM Research Almaden,
where he works on advanced information management systems that include
analytics, distributed ...


cid:1__=8FBB0B30DFE367208f9e8a93df938690918c8FB@ alok  singh ---09/22/2017
12:30:51 PM---Here are Niketan's question Thanks for taking time to answer
our questions and also for considering

From: alok singh <singh_alok@hotmail.com>
To: "dev@systemml.apache.org" <dev@systemml.apache.org>, "deron@apache.org"
<deron@apache.org>
Date: 09/22/2017 12:30 PM
Subject: Re: [PROPOSAL] R4ML Integration with SystemML





Here are Niketan's question

Thanks for taking time to answer our questions and also for considering to
help SystemML community. I have couple more questions:

Niketan:1.
In case there is inconsistency, do you (as R4ML developers) feel
comfortable changing R4ML interface to be compatible with our other APIs ?
May be you can go over the below two links and imagine adding a
corresponding R tab:
- MLContext Programming guide:
https://urldefense.proofpoint.com/v2/url?u=http-3A__apache.github.io_systemml_spark-2Dmlcontext-2Dprogramming-2Dguide&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=xyErlMsfwKjn_qfkXHpjLG8E1B70N5zVX-OWl5LU-yU&e= 


apache.github.io<https://urldefense.proofpoint.com/v2/url?u=http-3A__apache.github.io_systemml_spark-2Dmlcontext-2Dprogramming-2Dguide&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=xyErlMsfwKjn_qfkXHpjLG8E1B70N5zVX-OWl5LU-yU&e= >

apache.github.io
Spark MLContext Programming Guide. Overview; Spark Shell Example. Start
Spark Shell with SystemML; Create MLContext; Hello World; LeNet on MNIST
Example; DataFrame ...



- Algorithm wrappers:
https://urldefense.proofpoint.com/v2/url?u=http-3A__apache.github.io_systemml_algorithms-2Dclassification.html-23multinomial-2Dlogistic-2Dregression&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=TpQy-5v3cbhFJfGbEodsNvhrU8gDWexYBwN9x2eXzlc&e= 


ALOK: Hi Niketan

As pointed out earlier, R4ML is not just R interface
so it is based on the earlier product of IBM on R and it has many product
feature.

Also note that the pure ML Ctx and the cmd options for dml is not ideally
allow all the things user want to do in his ML code.
The solution could be to create wrapper to make user happy  . but we have
created those wrapper but those are in R and from user point for view it
feels that are just writing the R code

see some of the examples at

https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_SparkTC_r4ml_tree_master_R4ML_inst_examples&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=r4-fcsboHpxlbVf6KyY7C6ptdLcjmyT2g1hBHuqRa2s&e= 

https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_SparkTC_r4ml_blob_master_R4ML_inst_examples_r4ml.demo.mlogit.R&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=ScIkMGbMLlKu7VgjnDI5pDia2L8C3W_9fZXwZBjb7BI&e= 


NOTE: that R4ML uses combination of SparkR and DML and R to make user
experience best.

If the ultimate goal is to have just MLCtx based R interface than I think
it undermines and R4ML value proposition.
(We can definitely just expose MLCtx api. However calling Logistic
Regression example just for the purpose of MLCtx won't be best) R4ML.mlogit
has better apis

2. Classification - GitHub
Pages<https://urldefense.proofpoint.com/v2/url?u=http-3A__apache.github.io_systemml_algorithms-2Dclassification.html-23multinomial-2Dlogistic-2Dregression&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=TpQy-5v3cbhFJfGbEodsNvhrU8gDWexYBwN9x2eXzlc&e= >

apache.github.io
SystemML Algorithms Reference 2. Classification 2.1. Multinomial Logistic
Regression Description. The MultiLogReg.dml script performs both binomial
and multinomial ...




Niketan: 2. Other than providing R interface to SystemML as the above APIs,
what additional features/code R4ML plans to add in SystemML ? Just like we
want the R API to be functionally complete with our Python and Scala API,
we want Python and Scala APIs to  be functionally complete with the R API.
So a discussion on supporting the additional features in Python and Scala
APIs is required :)

ALOK: as talked in point 1) I think it will require a lot of work for scala
and python api to be in sync with r4ml api.
Also I feel that if the goal is too have just python, scala than we have to
do the coding at R4ML.

but I think goals was to merge this project.

I think @Fred if he can comment also that would be nice

Thanks
Alok



From: alok singh <singh_alok@hotmail.com>
Sent: Thursday, September 21, 2017 7:32 PM
To: dev@systemml.apache.org; deron@apache.org
Subject: Re: [PROPOSAL] R4ML Integration with SystemML

Hi

 We (me and Brendan) has been focusing on other things  like journeys apart
from new MLCtx changes. R4ML commits and PR you can also review,
I think code will definitely be maintained.

Alok





From: Deron Eriksson <deroneriksson@gmail.com>
Sent: Thursday, September 21, 2017 6:03 PM
To: dev@systemml.apache.org
Subject: Re: [PROPOSAL] R4ML Integration with SystemML

>
> * Looking over the github repo, apparently R4ML is not under active
> development/maintenance anymore (last commit Jul 20). So who would be
> willing to maintain and extend it?
>
> ALOK: We will doing development into it . there are open PR already.
>
>
No commits since Jul 20 does raise warning flags, as Matthias pointed out.
For some perspective, SystemML has 1013 commits in the last year (~2.78 per
day). No R4ML commits in 2 months is concerning for obvious reasons. It
implies no real work has been done on the project for months.




> * Providing wrappers for our algorithm scripts would be just a start
> because it hides our core value proposition of custom large-scale ML.
> Hence, we would also need an MLContext equivalent that allows to execute
> arbitrary DML scripts or R functions. Is there already a tentative design
> of such an API and if not, who would like to take it over?
>
> ALOK: Currently no out of box MLCtx.
>
>
I believe this also raises some warning flags. Looking over the code at
https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_SparkTC_r4ml_blob_master_R4ML_R_sysml.bridge.R&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=5kDETV7oPDlZ3OUDHX3lkMp6VxEJB9dUWCX7bZ1c76o&e= ,
  it looks


https://urldefense.proofpoint.com/v2/url?u=https-3A__avatars2.githubusercontent.com_u_13631156-3Fv-3D4-26s-3D400&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=YbUfZ7ntWQKbF6sqdbPrpVyZpRnB5ZwvnabMDRSyrw0&e= 


SparkTC/r4ml
github.com
r4ml - Scalable R for Machine Learning

like the code in the R4ML master branch utilizes an old API that does not
currently exist in SystemML. As Matthias pointed out, a key value
proposition of SystemML is customizable machine learning, which would
require an API that currently exists in the project.

That said, I believe an R API interface to SystemML is extremely valuable
and I think the whole SystemML community would benefit from the R API, and
I hope you will pursue the issue further. It looks like it has been in
development since June
(https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_SparkTC_r4ml_pull_50&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=fw5g1aTmnaaxg3-r142R9vfQbpvKlAPZPYbqHMe5Y-4&e= ).



https://urldefense.proofpoint.com/v2/url?u=https-3A__avatars2.githubusercontent.com_u_12959246-3Fv-3D4-26s-3D400&d=DwIFAw&c=jf_iaSHvJObTbx-siA1ZOg&r=HzVC6v79boGYQrpc383_Kao_6a6SaOkZrfiSrYZVby0&m=d7aHl15rr92bxoHo26sphduc7Q_4C0GizrRv_AR5pEM&s=Z7RXGGwxwpayjbVxUMlwBw1v-s03TDqZDeIlo496ITo&e= 


[WIP][I-50][R4ML-123] new MLContext API by aloknsingh · Pull Request #50 ·
SparkTC/r4ml
github.com
Developer's Certificate of Origin 1.1 By making a contribution to this
project, I certify that: (a) The contribution was created in whole or in
part by me and I have the right to subm...


Deron








Mime
  • Unnamed multipart/related (inline, None, 0 bytes)
View raw message