Manage inference packages about pystruct

[zaxtax]

I'm for all of these except for scripts to fetch other solvers, unless its a build step.

[vene]

also TODO: cython wrappers for libDAI

[amueller]

I added a small script to filter out not-installed inference methods in the meantime. @fgregg sent a PR to pyqpbo to make it installable via pip, I'm having a look right now.

Btw, I also implemented message passing in #13. It needs some heavy benchmarking, probably cythonizing.
Not sure if I should go with pointer pushing or typed memory views....

[fgregg]

So, @vene has been working on dai do we also need to include opengm? https://github.com/pystruct/pystruct/blob/master/pystruct/inference/inference_methods.py

[fgregg]

Also, we may want to not require that the user download all these methods. We could recommend one in the requirement.txt file, and the have the rest of the methods in requirement-test.txt file.

[amueller]

Totally. This was my intention in writing the helper function for the tests.
I think we stick with ad3 and recommend the rest.

It would be nice to have opengm pip installable. we should ask @DerThorsten about it.
All pystruct/opengm code is in inference_ogm, so if we can install Thorsten's branch via pip, we are golden.

[amueller]

opengm has a much broader and up-to-date collection of algorithms than libdai and also already includes python bindings.
That kind of begs the question why we want dai wrappers ;) But I guess LibDAI is pretty established while OpenGM is quite new - actually OpenGM can use LibDAI btw!

[fgregg]

So, should the default method be ad3 instead of lp?

On Mon, Jul 15, 2013 at 3:47 PM, Andreas Mueller
[email protected]:

opengm has a much broader and up-to-date collection of algorithms than
libdai and also already includes python bindings.
That kind of begs the question why we want dai wrappers ;) But I guess
LibDAI is pretty established while OpenGM is quite new - actually OpenGM
can use LibDAI btw!

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/7#issuecomment-21002400
.

773.888.2718
2231 N. Monticello Ave
Chicago, IL 60647

[amueller]

I think so. And I think @vene and @zaxtax agreed.

[amueller]

I was wondering if we should check whether ad3 is installed and fall back on lp
or just assume ad3 is installed.

[fgregg]

I think we should make it a requirement unless we find that people are
having a hard time installing it. In the meantime, we may want to
centralize all these 'default settings' [1] to a config file.

64bc339

On Mon, Jul 15, 2013 at 3:57 PM, Andreas Mueller
[email protected]:

I was wondering if we should check whether ad3 is installed and fall back
on lp
or just assume ad3 is installed.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/7#issuecomment-21003144
.

773.888.2718
2231 N. Monticello Ave
Chicago, IL 60647

[amueller]

On 07/15/2013 11:03 PM, Forest Gregg wrote:

I think we should make it a requirement unless we find that people are
having a hard time installing it.
Fine with me.
In the meantime, we may want to
centralize all these 'default settings' [1] to a config file.
We definitely need to centralize them.
I'm not sure I like the idea of a config file. Why not in the parent class?

I guess the main question is whether we want to allow the user to change
the default behavior.
I have absolutely no experience with that (aka sklearn doesn't do it).
It might make sense here.
That means we have to look for the config-file in user-space, right, like
~/.pystructrc
or something like that?

[zaxtax]

Well numpy as a site.cfg file in their top directory, https://github.com/numpy/numpy/blob/master/site.cfg.example

[amueller]

I don't see a benefit of having a config file instead of doing it in the source, if it is not in user-space.

[fgregg]

I think you are right, it makes more sense to set it in the StructuredModel
class.

Having a loose coupling with the inference methods make sense, since there
is so much fllux in the available libraries.

On Mon, Jul 15, 2013 at 4:09 PM, Andreas Mueller
[email protected]:

On 07/15/2013 11:03 PM, Forest Gregg wrote:

I think we should make it a requirement unless we find that people are
having a hard time installing it.
Fine with me.

In the meantime, we may want to
centralize all these 'default settings' [1] to a config file.
We definitely need to centralize them.
I'm not sure I like the idea of a config file. Why not in the parent class?

I guess the main question is whether we want to allow the user to change
the default behavior.
I have absolutely no experience with that (aka sklearn doesn't do it).
It might make sense here.
That means we have to look for the config-file in user-space, right, like
~/.pystructrc
or something like that?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/7#issuecomment-21003980
.

773.888.2718
2231 N. Monticello Ave
Chicago, IL 60647

[amueller]

the numpy config seems to be for building, right?

[zaxtax]

Right.

On Mon, Jul 15, 2013 at 2:17 PM, Andreas Mueller
[email protected]:

the numpy config seems to be for building, right?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/7#issuecomment-21004542
.

[amueller]

How about we try to do it in the base class and if

• it is annoying to implement or
• we find it annoying to overwrite it with our favorite every time

we introduce a config file?

[fgregg]

👍