[Kwiver-users] Fwd: [Kwiver-announce] Best practices for passing inherited types through Sprokit Pipeline

Linus Sherrill linus.sherrill at kitware.com
Wed Nov 15 15:39:02 EST 2017 

Processes like the muxer implement a method that is called when the
connection is attempted. It then looks at the requested type for the port
and can dynamically create a port with that data type string. The process
also needs to deal with the actual data type that is being passed. These
flow control processes (e.g. mux, collate) just pass the input data item to
an output port without actually extracting it from the boost::any in the
data packet.

I am working on an idea to support casting data elements using the port
trait infrastructure. The call would look something like

auto value = get_from_port_using_trait<desired-type>( port-name );

This would extract the value from the port based on the type specified in
the trait and then attempt to cast it to the desired type. Will need to
determine how to handle errors when the cast fails. This approach does not
address how to deal with the logical data type string associated with the
port.

One approach would be to specific type strings that refer to the base types
(e.g. track_set) and then expect the processes to use above call to convert
to desired type. Presumably this situation would not be all that common
that a polymorphic type would be passed.

Anyhow, it is just a concept in the early stages...



On Wed, Nov 15, 2017 at 2:58 PM, Matthew Leotta <matt.leotta at kitware.com>
wrote:

> I don’t think we should.  The right way to handle this (in your case) is
> to pass through the input port type to the output.  I believe I remember
> seeing examples of this in sprokit for more generic processes, like a muxer
> and demuxer.  Those processes say “I don’t care what the input type is, but
> the output type needs to match whatever the input type was”.  What I don’t
> know is whether we do both the input/output type matching and the specify
> type requirements at the same time.
>
> My understanding of the state of sprokit is a little outdated so someone
> correct me if I’m wrong about the type matching thing.
>
> —Matt
>
> On Nov 15, 2017, at 2:51 PM, Brian Clipp <brian.clipp at kitware.com> wrote:
>
> Would we allow down-casting?
>
> -B
>
> On Wed, Nov 15, 2017 at 2:51 PM Matthew Leotta <matt.leotta at kitware.com>
> wrote:
>
>> I would prefer a string the makes the derivation order explicit rather
>> than just an arbitrary list, like “track_set.feature”, but I could be
>> convinced otherwise.  This way you can just match substrings from the front
>> until you find a match.
>>
>>
>> On Nov 15, 2017, at 2:37 PM, Matt Phillips <matt.phillips at kitware.com>
>> wrote:
>>
>> Forget the exact signature but
>>
>> class feature_track_set : public ...
>> {
>>    ...
>>    static std::string type_string() { return
>> "feature_track_set,track_set"; }
>> }
>>
>> Is one way to do it, just turn the name into a list of (permissible)
>> names.  This creates coupling issues but there can be an early runtime
>> check to ensure that given "A,B", B is actually a base of A.
>>
>> On Wed, Nov 15, 2017 at 2:25 PM, Matthew Leotta <matt.leotta at kitware.com>
>> wrote:
>>
>>> Yes, exactly.  That what I was getting at with the earlier idea of
>>> substring processing on the type strings, but there is probably a better
>>> way to implement it than that.
>>>
>>> On Nov 15, 2017, at 2:23 PM, Matt Phillips <matt.phillips at kitware.com>
>>> wrote:
>>>
>>> So this new idea is simpler but more permissive than it needs to be.
>>> feature_track_set is a track_set but isn't a object_track_set.  Only the
>>> former would be allowed, so Sprokit's type-checking system would have to be
>>> made to understand the inheritance relationships somehow, but what you just
>>> described Brian would be prohibited.
>>>
>>> Matt
>>>
>>> On Wed, Nov 15, 2017 at 2:15 PM, Brian Clipp <brian.clipp at kitware.com>
>>> wrote:
>>>
>>>> Yeah, I'm not a fan of calling every flavor of track_set a track_set.
>>>> That could allow us to connect object_track_set ports to feature_track_set
>>>> ports.  We could catch that when constructing the graph.
>>>>
>>>> -B
>>>>
>>>>
>>>> On Wed, Nov 15, 2017 at 2:10 PM Matthew Leotta <matt.leotta at kitware.com>
>>>> wrote:
>>>>
>>>>> I’d rather not tell sprokit to ignore type checking on some subset of
>>>>> types, but not others.  However, equivalently you could define all
>>>>> track_set types to be “track_set” regardless of the specific derived
>>>>> class.  Then each process would need to check at run time if the track_set
>>>>> it got can be down cast to the type it was expecting.  That approach is
>>>>> equivalent, but probably less of a hack.  It’s still not ideal.
>>>>>
>>>>> —Matt
>>>>>
>>>>> On Nov 15, 2017, at 2:01 PM, Brian Clipp <brian.clipp at kitware.com>
>>>>> wrote:
>>>>>
>>>>> Agreed.  Matt P. had an interesting suggestion.  What if Sprokit's
>>>>> graph type checking would allow some potentially wrong graphs through to
>>>>> allow for up-casting?  It could catch really egregious issues, e.g. int
>>>>> passed into a port expecting a string, but not subtle issues like a
>>>>> track_set passed into a feature_track set port where the track set is in
>>>>> fact an object_track_set.
>>>>> Basically, the type checker would consider up-casts and down-casts as
>>>>> being equivalent types.  Most errors would be caught at pipeline
>>>>> construction time but a few might only be caught at run-time.
>>>>>
>>>>> -Brian
>>>>>
>>>>>
>>>>> On Wed, Nov 15, 2017 at 1:53 PM Matthew Leotta <
>>>>> matt.leotta at kitware.com> wrote:
>>>>>
>>>>>> This has been a long standing issue with sprokit, from what I
>>>>>> recall.  It seems to me that we need a mechanism for allowing sprokit type
>>>>>> checking to allow for casting up the inheritance hierarchy.  A
>>>>>> feature_track_set is a track_set, so any process that can process a
>>>>>> track_set should also be able to process a feature_track_set.  I’m not sure
>>>>>> if there is an easy way of checking this because I think sprokit just does
>>>>>> string name comparison for type checking.  Maybe the type name strings
>>>>>> could be “track_set”, “track_set.feature”, “track_set.object” and we could
>>>>>> do substring comparison?  Anyway, that is probably a discussion for another
>>>>>> day.  In your case, the issue is more challenging because there is also
>>>>>> down casting involved.  Ideally you want to encode that the input and
>>>>>> output types for this process must be 1) the same exact type and 2) a
>>>>>> track_set or some derivative of it.  I’m not sure that we can do that in
>>>>>> sprokit yet.
>>>>>>
>>>>>> I mostly agree with the best practices rules in the PPTX, except that
>>>>>> I’d like to see sprokit improved so that the first rule is not needed.  For
>>>>>> now, I think your Option 2 (slide 4) makes the most sense.
>>>>>>
>>>>>> —Matt
>>>>>>
>>>>>> On Nov 15, 2017, at 1:22 PM, Brian Clipp <brian.clipp at kitware.com>
>>>>>> wrote:
>>>>>>
>>>>>> Pushing the tread over to kwiver-users.  I attached the original ppt.
>>>>>>
>>>>>> -Brian
>>>>>>
>>>>>> ---------- Forwarded message ---------
>>>>>> From: Matthew Leotta <matt.leotta at kitware.com>
>>>>>> Date: Wed, Nov 15, 2017 at 1:08 PM
>>>>>> Subject: Re: [Kwiver-announce] Best practices for passing inherited
>>>>>> types through Sprokit Pipeline
>>>>>> To: Matt Phillips <matt.phillips at kitware.com>, Brian Clipp <
>>>>>> brian.clipp at kitware.com>
>>>>>> Cc: <kwiver-announce at public.kitware.com>
>>>>>>
>>>>>>
>>>>>> Guys,
>>>>>>
>>>>>> Let’s take this discussion to the kwiver-users list rather than
>>>>>> kwiver-announce.  kwiver-annouce is supposed to be for major announcements,
>>>>>> like new versions.
>>>>>>
>>>>>> Of course this make me wonder if we really need kwiver-announce
>>>>>> anymore.  Is there anyone on this mailing list who is not also on
>>>>>> kwiver-users and who only wants to receive very infrequent notifications
>>>>>> about KWIVER releases and other major announcements?  If this describes
>>>>>> you, please send me a response privately.  If I get no responses in a week
>>>>>> or so, I think we should retire this list and use only kwiver-users.  That
>>>>>> pattern aligns better with what our other open source projects do for
>>>>>> mailing lists.
>>>>>>
>>>>>> Thanks,
>>>>>> Matt
>>>>>>
>>>>>>
>>>>>> On Nov 15, 2017, at 12:56 PM, Matt Phillips <
>>>>>> matt.phillips at kitware.com> wrote:
>>>>>>
>>>>>> Is the issue here just slicing when derived class objects are passed
>>>>>> to Keyframe selector?  I'm not familiar off the top of my head with how
>>>>>> sprokit does serialization/deserialization but might making that
>>>>>> 'slice-proof' be the answer?  By e.g. storing type information and using
>>>>>> factory methods at construction.
>>>>>>
>>>>>> Matt
>>>>>>
>>>>>> On Wed, Nov 15, 2017 at 12:46 PM, Brian Clipp <
>>>>>> brian.clipp at kitware.com> wrote:
>>>>>>
>>>>>>> Hi All.
>>>>>>>
>>>>>>> I ran into a case yesterday where I had a process that output a
>>>>>>> child class and a process that expected an input from it of the parent
>>>>>>> class.  Sprokit's type checking fails in that case.  It lead me to spend
>>>>>>> some time thinking about what types to pass between processes.  After some
>>>>>>> discussion with Aaron and Matt B. these are some suggestions I've arrived
>>>>>>> at.
>>>>>>>
>>>>>>> Thoughts?
>>>>>>>
>>>>>>> Brian
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> Kwiver-announce mailing list
>>>>>>> Kwiver-announce at public.kitware.com
>>>>>>> http://public.kitware.com/mailman/listinfo/kwiver-announce
>>>>>>>
>>>>>>>
>>>>>> _______________________________________________
>>>>>> Kwiver-announce mailing list
>>>>>> Kwiver-announce at public.kitware.com
>>>>>> http://public.kitware.com/mailman/listinfo/kwiver-announce
>>>>>>
>>>>>>
>>>>>> <Inherited Types in Sprokit Pipelines.pptx>_______________
>>>>>> ________________________________
>>>>>> Kwiver-users mailing list
>>>>>> Kwiver-users at public.kitware.com
>>>>>> http://public.kitware.com/mailman/listinfo/kwiver-users
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>> _______________________________________________
>>>> Kwiver-users mailing list
>>>> Kwiver-users at public.kitware.com
>>>> http://public.kitware.com/mailman/listinfo/kwiver-users
>>>>
>>>>
>>>
>>>
>>
>>
>
> _______________________________________________
> Kwiver-users mailing list
> Kwiver-users at public.kitware.com
> http://public.kitware.com/mailman/listinfo/kwiver-users
>
>


-- 
*Linus Sherrill - *Staff R&D Engineer
Kitware, Inc.
28 Corporate Drive
Clifton Park, NY 12065-8662
E: linus.sherrill at kitware.com
P: 518.881.4400
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/kwiver-users/attachments/20171115/ad80f3d8/attachment-0001.html>

More information about the Kwiver-users mailing list