Regarding recommenders, drivers, and import/export:

I’ve got Sebastian’s cooccurrence code wrapped with a driver that reads text 
delimited files into a drm for use with cooccurrence. Then it writes the 
indicator matrix(es) as text delimited files with user specified IDs. It also 
has a proposed Driver base class, Scala based option parser and 
ReadStore/WriteStore traits. The CLI will be mostly a superset of the 
itemsimilarity in legacy mr. The read/write stuff is meant to be pretty generic 
so I was planning to do a DB and maybe JSON example (some day). There is still 
a bit of functional programming refactoring and the docs are not up to date.

With cooccurrence working we could do something that replaces all the 
cooccurrence  recommenders (in-memory and MR) with one codebase. Add Solr and 
you have a single machine server based recommender that we can supply with an 
API similar to the legacy in-memory recommender. The cool thing is that It will 
scale out to a cluster with Solr and HDFS, requiring only config changes. The 
downside is that it requires at least a standalone local version of Spark to do 
the cooccurrence. BTW this would give us something people have been asking 
for—a recommender service.

Is anyone else interested in CLI, drivers, read/write in the import/export 
sense? Or a new architecture for the recommenders? If so, maybe a separate 
thread?

On May 29, 2014, at 7:03 AM, Ted Dunning <[email protected]> wrote:

Andrew,

Sebastian and I were talking yesterday and guessing that you would be
interested in this soon.  Glad to know the world is as expected.

Yes. This needs to happen at least at a very conceptual level.  For
instance, for classifiers, I think that we need to have something like:

   - progressively train against a batch of data
        questions: should this do multiple epochs?  Throw an exception if
on-line training not supported?  throw an exception if too little data
provided?

   - classify a batch of data

   - serialize a model

   - de-serialize a model

Note that a batch listed above should be either a bunch of observations or
just one.

Question: does this handle the following cases:

- naive bayes
- SGD trained on continuous data
- batch trained <mumble> classifiers
- downpour type classifier training

?



On Wed, May 28, 2014 at 6:25 PM, Andrew Palumbo <[email protected]> wrote:

> This may be somewhat tangential to this thread, but would now be a good
> time to start laying out some scala traits for
> Classifiers/Clusterers/Recommenders?  I am totally scala-naive, but have
> been trying to keep up with the discussions.
> 
> I don't know if this is premature but it seems that now that the DSL data
> structures have been at least sketched out if not fully implemented,  it
> would be useful to have these in place before people start porting too much
> over.  It might be helpful in bringing in new contributions as well.
> 
> It could also help regarding people's questions of integrating a future
> wrapper layer.
> 
> 
> 
>> From: [email protected]
>> Date: Wed, 28 May 2014 17:10:43 -0700
>> Subject: Re: do we really need scala still
>> To: [email protected]
>> 
>> +1
>> 
>> Let's use a successful scala model as a suggestion about where to go.  It
>> seems plausible that Java could emulate the building of a lazy DSL
> logical
>> plan and then poke it in plausible ways with the addition of a wrapper
>> layer.  But that only helps if the Scala layer succeeds.
>> 
>> 
>> 
>> On Tue, May 27, 2014 at 10:56 AM, Dmitriy Lyubimov <[email protected]
>> wrote:
>> 
>>> Also, i think that this is leaning towards false dilemma fallacy.
> Scala and
>>> java models could happily exist at the same time and hopefully, minimal
>>> fragmentation of the project if done with precision and care.
>>> 
>>> 
>>> On Tue, May 27, 2014 at 10:46 AM, Dmitriy Lyubimov <[email protected]
>>>> wrote:
>>> 
>>>> 
>>>> not sure there's much sense in taking user survey if we can't act on
>>> this.
>>>> In our situation, unfortunately, we don't have that many ideas to
> choose
>>>> from, so there's not much wiggle room imo. It is more like
> reinforcement
>>>> learning -- stuff that doesn't get used or supported, just dies
> .that's
>>> it.
>>>> Scala bindings, though thumb up'd internally, are yet to earn this
> status
>>>> externally. In that sense we always have been watching for
> use/support,
>>>> that's why we culled out tons of stuff. Nothing changes going
> forward (at
>>>> least at this point). If we have tons of new ideas/contributions,
> then it
>>>> may be different. What is weak, dies on its own pretty evidently
> without
>>>> much extra effort.
>>>> 
>>>> 
>>>> On Tue, May 27, 2014 at 10:32 AM, Pat Ferrel <[email protected]>
>>> wrote:
>>>> 
>>>>> We are asking that anyone using Mahout as a lib or in the DSL-shell
> to
>>>>> learn Scala. While I still think it’s the right idea, user’s may
>>> disagree.
>>>>> We should probably either solicit comments or at least keep an eye
> on
>>>>> reactions to this. Spark took this route when the question was even
>>> more in
>>>>> doubt and so is at least partially supporting multiple bindings.
>>>>> 
>>>>> Not sure how far we want to carry this but we could supply Java
> bindings
>>>>> to the CLI-type things pretty easily.
>>>>> 
>>>>> 
>>>>> On May 26, 2014, at 2:43 PM, Dmitriy Lyubimov <[email protected]>
>>> wrote:
>>>>> 
>>>>> Well, first, functional programming in java8 is about 2-3 years
> late to
>>>>> the
>>>>> scene. So the reasoning along the lines, hey, we already are using
> tool
>>> A,
>>>>> and now tool B is available which is almost as good as A, so let's
>>> migrate
>>>>> to B, is fallible. Tool B must demonstrate not just matching
>>> capabilities,
>>>>> but far superb, to justify cost of such migration.
>>>>> 
>>>>> Second, as other pointed, java 8 doesn't really match scala, not yet
>>>>> anyway. One important feature of scala bindings work is proper
> operator
>>>>> overload (R-like DSL). That would not be possible to do in java 8,
> as it
>>>>> stands. Yes, as other pointed, it makes things concise, but most
>>>>> importantly, it also makes things operation-centric and eliminates
>>> nested
>>>>> calls pile-up.
>>>>> 
>>>>> Third, as it stands today, it would also presentn a problem from the
>>> Spark
>>>>> integration point of view. Spark does have java bindings, but first,
>>> they
>>>>> are underdefined (you can check spark list for tons of postings
> about
>>>>> missing equivalent capability), and they are certainly not
>>> java-8-vetted.
>>>>> So java api in Spark for java 8 purposes, as it stands, is a moot
> point.
>>>>> 
>>>>> There are also a number other goodies and clashes that exist -- use
> of
>>>>> scala collections vs. Java collections, clean functional type
> syntax,
>>>>> magic
>>>>> methods, partially defined functions, case class matchers,
> implicits,
>>> view
>>>>> and context bounds etc. Etc., all that sh$tload of acrobatics that
> comes
>>>>> actually very handy in existing  implemetations and has no
> substitute in
>>>>> Java 8.
>>>>> On May 25, 2014 12:48 PM, "bandi shankar" <[email protected]>
>>> wrote:
>>>>> 
>>>>>> Hi,
>>>>>> 
>>>>>> I was just thinking , do we still need scala . Since in java 8 we
> have
>>>>>> all(probably) kind of feature provided by scala.
>>>>>> Since I am new to group , so just thinking why not to make mahout
> away
>>>>>> from scala. Is there any specific reason to adopt scala.
>>>>>> 
>>>>>> Bandi
>>>>>> 
>>>>> 
>>>>> 
>>>> 
>>> 
> 
> 

Reply via email to