Re: [boost] [accumulators] Help understanding implementation of weighted accumulators

Subject: Re: [boost] [accumulators] Help understanding implementation of weighted accumulators
From: Gabriel Redner (gredner_at_[hidden])
Date: 2013-05-01 23:06:46

Hi Eric,

Thanks very much for your response.  That does help, and I have some follow-ups.

First, let me make sure I am understanding correctly.  If my Sample is
some FancyType whose default constructor does not do the right thing
(or perhaps does not exist), I should do:

accumulator_set<FancyType, features<some_features> > acc(sample =
some_initial_fancytype_value);

Is that right?  If so, I would say that:
- This usage is not mentioned in the docs anywhere, and
- It seems strange to overload the meaning of 'sample' in this way.
Wouldn't it be clearer and accomplish the same goal to use a different
name?  "initialValue" perhaps?
- Do all accumulators support this feature?  I can imagine some that
might not (indeed I am implementing one).  This is not mentioned in
the docs either.

Second, I have a sort-of-question-but-mostly-a-kvetch about defining
keywords for accumulators that need initialization parameters (e.g.
tail).  The library itself uses the macro
BOOST_PARAMETER_NESTED_KEYWORD for this, which is not documented.  I
do not understand what this macro does - why the name, why the alias,
etc?  My own accumulators need initialization parameters, and I have
just been aping what the library itself does - using the macro
although it's not documented, and hoping for the best.  It seems to be
working, but I'd like to understand if this is the right approach,
whether I can rely on this undocumented macro, and what it's doing.
Can you shed any light?

Thanks,
-Gabe

On Wed, May 1, 2013 at 8:23 PM, Eric Niebler <eniebler_at_[hidden]> wrote:
> On 5/1/2013 1:13 PM, Gabriel Redner wrote:
>> Hi folks,
>>
>> I am trying to implement a weighted accumulator, but I'm having
>> trouble understanding a few things about the existing ones.
>>
>> Looking at the implementation of weighted_sum, I see (weighted_sum.hpp:37):
>>
>>         template<typename Args>
>>         weighted_sum_impl(Args const &args)
>>           : weighted_sum_(
>>                 args[parameter::keyword<Tag>::get() | Sample()]
>>                   * numeric::one<Weight>::value
>>             )
>>         {
>>         }
>>
>> My questions are:
>>
>> - It's taking the first sample from the args and multiplying it by a
>> weight of 1.  Isn't this wrong - ignoring the real weight of this
>> sample?  Shouldn't this produce incorrect results?
>
> It's not taking the first sample. It's (optionally) taking an element
> representing zero. In most cases, a default-constructed Sample() is
> satisfactory for a zero element, but in some cases it's not. Consider
> what happens if your Sample type is std::vector<int>, and all your
> samples are going to be 3-element vectors. Then your zero is a
> std::vector<int> with 3 zeros. Any other starting value would cause all
> future computation to fail.
>
> And we multiply by one to make the type promotion come out right.
>
>> - I've never understood anyways how there can be an initial argument
>> during construction of the accumulator.  I haven't found any
>> explanation in the docs beyond "Maybe there is an initial value in the
>> argument pack".  Maybe?  How?
>
> I hope that clears it up.
>
> --
> Eric Niebler
> Boost.org
> http://www.boost.org
>
> _______________________________________________
> Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost