commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Phil Steitz <phil.ste...@gmail.com>
Subject Re: [math] getting changes included into commons-math (was Re: Home for the colt fork)
Date Thu, 10 Dec 2009 12:46:37 GMT
Jake Mannix wrote:
> On Wed, Dec 9, 2009 at 11:09 AM, Benson Margulies <bimargulies@gmail.com>wrote:
> 
>> This is interesting. We have a raft of mathematically qualified
>> committers on Mahout, and this message asking for help on
>> commons-math, and a raft of code marooned at mahout that wants to be
>> in commons math. If I were one of those mathematically competant
>> individuals, I'd be off attaching a patch or three to a JIRA or two
>>
> 
> The commons-math linear APIs have been described as effectively locked
> until 3.0, due to back-compat requirements.  This means that any code
> contributed
> into c-math would live in a parallel (no pun intended) to the linear
> primitives which
> exist already in there.
> 
> Adopting something like MTJ or Colt in Mahout turned out to be easier,
> because
> we are on release 0.2 (heading for 0.3 now), and have less stringent
> back-compat
> requirements, so we are overhauling our linear apis (read: even user-facing
> interface changes) to take advantage of useful parts of Colt, and are
> planning on
> using our Colt fork as the underlying implementation.
> 
> Commons-math expressed that changing linear APIs is not something they can
> do,
> given the maturity of their library, so where would Colt *go* in c-math?
> It's own
> submodule, having its own eigendecompositions and svd and so forth, running
> parallel to the current c-math impls?  Why?
> 
> Who would maintain it and write tests for it, and how do you explain to
> end-users which they should use?
> 
> 
> 
>> On Wed, Dec 9, 2009 at 1:48 PM, Luc Maisonobe <Luc.Maisonobe@free.fr>
>> wrote:
>>> Ted Dunning a écrit :
>>>> Actually, the reason that we have Colt in Mahout is it has proven
>> impossible
>>>> to get changes into commons math.  We really, really wanted to use
>> commons
>>>> math rather than have our own linear algebra package, but it just proved
>>>> impossible and we didn't want to wait forever.
>>> If you really, really wants to use commons math and want changes to be
>>> included, contribute them.
> 
> I have submitted patches for the following tickets: MATH-312 (and acceptance
> of that patch blocks my patch for MATH-314), MATH-316 and MATH-317, none
> of which have appear to have had much progress on.  All of my patches come
> with unit tests for new functionality.
> 
> On the other hand, when I opened the discussion about extending the
> functions
> package to enable composable functions (MATH-313), I got an entirely hostile
> response, which only tempered as far as "+0" on adding it after discussion.
> 
> In particular, my first step at making commons-math something Mahout could
> standardize on for linear work was MATH-312, which I did submit a patch for,
> and revised it many times after discussion about what is acceptable practice
> in c-math.  Not yet applied, months later.  It's probably far out of date
> now...
> 
> Similarly, when I tried to ask what the status on decisions on whether to
> adopt
> MTJ or Colt, the statement by Phil was basically that commons-math would not
> adopt anything which had any external dependencies or
> not-easily-human-readable java source (which ruled out MTJ because of f2j
> produced code), and which had to be fully tested and maintained prior to
> adoption (which rules out Colt which has no unit tests yet).

We agreed early on that commons-math was going to be self-contained.
  It is fine to reopen that discussion.  I stated my opinion, which
is to stay away from *required* external dependencies.  I have been
wrong before, and will be wrong again.  Clear arguments can
enlighten me.  I also feel obligated to support the code that we
ship.  Others may disagree with this and feel that it is OK to ship
code that we - the committers voting on the release - cannot debug
or understand.  It will take some very "enlightening" arguments to
get me to agree to releasing code that I can't understand.
Regarding unit tests, the answer is simple - volunteer to write them
and include them in patches.  Code without unit tests is not
complete code. If I commit it, it means that *I* am going to have to
write the unit tests and I am going to do that before committing.
Patches without unit tests will therefore take longer to commit.

Phil
> 
> Ted and I weren't making "requests" for other people to do work, we were
> wondering whether even offers to do some of the work would be accepted,
> and for many of the questions/suggestions we had, it seems the desires
> and requirements of the Mahout community were incompatible with those
> of commons-math.
> 
>   -jake
> 
> 
>>  > I think the only change that was proposed and not done because of lack
>>> of consensus was the inclusion of MTJ (and I don't consider the
>>> discussion closed on that topic either, so it may still happen some
>>> day). All the other changes that are desired are simply lacking someone
>>> to do the work. There were proposals to extend the linear algebra API,
>>> proposals to add more support for sparse matrices, proposals to get
>>> partial decomposition ... But sparse contributions (pun intended).
>>>
>>> I try to do what I can, but as you have probably seen have been rather
>>> silent since 2.0 release. For my part, I really, really need help. I
>>> would like to fix the problems in the eigen decomposition and SVD but
>>> need a good kick to get on it, and having only requests and no help is
>>> not really motivating.
>>>
>>> Luc
>>>
>>>> If that problem were solved, then it would be great to depend on commons
>>>> math.  If that problem isn't solved, then there is no way to depend on
>>>> commons math.
>>>>
>>>> On Tue, Dec 8, 2009 at 6:19 AM, Benson Margulies <bimargulies@gmail.com
>>> wrote:
>>>>> We can't possibly have a dependency on Mahout in the long term. Either
>>>>> we all go shares on code in some other piece of commons, or we end up
>>>>> with two forks, which would be sad.
>>>>>
>>>>> On Tue, Dec 8, 2009 at 8:33 AM, James Carman <
>> james@carmanconsulting.com>
>>>>> wrote:
>>>>>> I wouldn't like to see a dependency on mahout code in a "commons"
>>>>>> library.  That seems kind of backwards.  If Mahout wants to offload
>>>>>> this stuff, we can move it into a library in commons (which is
>>>>>> typically how stuff used to happen in Jakarta).
>>>>>>
>>>>>> On Tue, Dec 8, 2009 at 8:18 AM, Benson Margulies <
>> bimargulies@gmail.com>
>>>>> wrote:
>>>>>>> Mahout now has a fork of a portion of the 'category A' portion
of the
>>>>>>> CERN colt library forked. The Mahout fork is, of course, in the
>> Mahout
>>>>>>> tree under a Mahout Java package and Maven triple.
>>>>>>>
>>>>>>> I want to use the collections classes from Mahout as the core
to a
>> new
>>>>>>> set of commons-primitives classes that do the useful things that
GNU
>>>>>>> Trove does.
>>>>>>>
>>>>>>> The classes I want to start from depend on the classes that are
in
>> the
>>>>>>> Mahout fork.
>>>>>>>
>>>>>>> As a temporary expedient, I can depend on their there. However,
I
>>>>>>> submit that it would be more better if the mathematical code
were in
>>>>>>> commons-math. Was this option explored?
>>>>>>>
>>>>>>> ---------------------------------------------------------------------
>>>>>>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>>>>>>> For additional commands, e-mail: dev-help@commons.apache.org
>>>>>>>
>>>>>>>
>>>>>> ---------------------------------------------------------------------
>>>>>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>>>>>> For additional commands, e-mail: dev-help@commons.apache.org
>>>>>>
>>>>>>
>>>>> ---------------------------------------------------------------------
>>>>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>>>>> For additional commands, e-mail: dev-help@commons.apache.org
>>>>>
>>>>>
>>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>>> For additional commands, e-mail: dev-help@commons.apache.org
>>>
>>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>> For additional commands, e-mail: dev-help@commons.apache.org
>>
>>
> 


---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org


Mime
View raw message