[mlpack] mlpack video meeting, Wednesday Aug. 28th, 1400 UTC
Ryan Curtin
ryan at ratml.org
Wed Aug 28 12:21:50 EDT 2019
Hey everyone,
Thanks to those who attended.
Here are the slides that I presented on what has been going on with
mlpack:
http://www.ratml.org/misc/mlpack-meeting-slides.pdf
Below are my notes for the meeting. It's long, but the notes are pretty
comprehensive and might involve you, so please, take a look. :)
====
During the meeting we did some release planning. Rahul pointed out that
PRs #1931 and #1941 make some reverse-incompatible changes to the
reinforcement learning API; specifically, the Action structs are changed
to hold a 'double' instead of a 'double[]'. We are still discussing
whether this is part of the "public" or "internal" API, and the decision
will determine whether the release is 4.0.0 or 3.2.0.
We can release the next version quite soon (within a week or two). We
discussed a number of different PRs to see what we can wait on and what
we won't:
- Linear SVM CLI and Python bindings
https://github.com/mlpack/mlpack/pull/1935
* This seems close to approval, but there is still one detail about
parallel SGD to work out, so it may not be ready in time. I don't
think we should wait for it, since we can just have another
release when that is ready.
- Reclaim unused KDE tolerance
https://github.com/mlpack/mlpack/pull/1984
* This should be merged in the next day or two.
- Julia bindings and Go bindings
https://github.com/mlpack/mlpack/pull/1949
https://github.com/mlpack/mlpack/pull/1884
* These seem ready, but finding time to review them is hard, so
maybe we will not be able to merge them for a little while. I
don't see a need to wait on these.
- Add getter methods to layers
https://github.com/mlpack/mlpack/pull/1985
* This may take a little while---no need to wait on it.
- Create the header to determine the layer type
https://github.com/mlpack/mlpack/pull/1987
* This can probably be merged in the next couple days, so we can
wait.
- Improve TransposedConvolution layer
https://github.com/mlpack/mlpack/pull/1493
* This is ready, but needs a review still. Hopefully we can get a
review done in the next couple days.
Some PRs we are not sure about; will they be ready soon?
if you know about these, do you have any input? I'll send a ping to
those involved in the next few days if I don't hear anything. :)
- String encoding bindings (appears approved, is it ready to merge?)
https://github.com/mlpack/mlpack/pull/1980
- MiniBatchDiscrimination layer
https://github.com/mlpack/mlpack/pull/1913
- Generator update for GANs
https://github.com/mlpack/mlpack/pull/1933
If there are any other PRs that weren't mentioned in this list that you
think should be merged before this release, respond to this email and
point them out---we can get it figured out. :)
====
We also talked about the different GSoC projects. Most of the summaries
can be found on the blog; each is linked to in the meeting slides. Lots
of very cool work happened (and is continuing to happen) this summer!
====
Then we had some more open discussion about the direction of the
project, the ways that we communicate, and the ways that we work
together. Here are some of the comments that were made:
- Suryo pointed out that it would be a great idea to have some
platforms that can help people connect better. For instance, the
website is a nice step forward, but perhaps there could be some ways
that the contributors can open up and talk about other things, like
forum threads or something like this, where every individual topic
and idea can be documented. This is not just about mlpack and
ensmallen, and could be wider scope.
- It could be nice to make some ensmallen stickers. I'll see if I can
do this in the upcoming weeks.
- Toshal asked about Bandicoot; this is still a work in progress, so
it's not ready yet, but I'm hopeful that in the upcoming weeks or
months that the groundwork can be laid in place so that we can see
mlpack run on GPUs, even in a prototype form.
- Suryo and Rahul discussed combinatorial optimization, which is pretty
fundamental in CS (see, e.g., the traveling salesman problem).
ensmallen could be improved to handle these problems. Right now it
has categorical feature support, and maybe this could be used as a
start, but in any case there are improvements that could be made and
additional optimizers that could be added. They will follow up with
each other through email or IRC to discuss this further.
- Suryo had a great comment about discoverability and documentation.
When we release mlpack and other parts of the system, it can be
really helpful to have some blog posts that discuss the improvements
and how they can be used. Basically, a lot of people are not looking
for documentation but instead stories and guidance on how they can
apply the code that we have to their specific problems. So, it would
be great if we could do more of that, and I agree---how much our code
gets used depends on how accessible we can make it to the public, and
blog posts or examples or stories are great ways to do that.
- In addition, Rahul suggested that we assemble a blog post for the
next release, to highlight all the new features.
I think that all of these are great ideas! But personally, from my end
I don't have time to execute all of them (or even many of them). So if
you feel inspired by what you read here, help make it a reality! For
the next release, I think a blog post would be an awesome idea, but I
don't have time (or actually knowledge of all the relevant code) to
write up all of these sections of a blog post. So if you would like to
write something about your feature or some other features that you know
got included, it would be awesome, and could help contribute to a nice
blog post with the release. Send an email or a reply to this or an IRC
message and let's make it happen!
====
In the end, we decided that we should do these video meetings more
often, but just like the above, the key issue is finding the time. :)
If someone else wants to put together an update presentation or agenda
or help make them happen, it can make it a lot easier for them to happen
more often.
Hopefully, we can have another meeting in the next month or two months.
====
I hope these notes are useful! Please feel free to respond to any of
the comments here. The meeting itself is just a starting place for
discussion; it's not final. :)
On Wed, Aug 28, 2019 at 09:09:55AM -0400, Ryan Curtin wrote:
> Hey everyone,
>
> Just a friendly reminder---if you'd like to join today's video meeting,
> it will be in about an hour, at the Zoom link below. If we have
> problems, I guess we can try to figure out what to do in IRC. :)
>
> Thanks!
>
> Ryan
>
> On Sun, Aug 25, 2019 at 09:20:37PM -0400, Ryan Curtin wrote:
> > Hello everyone,
> >
> > I meant to send this out on Friday but unfortunately got sidetracked!
> > Anyway, I took a look at the whenisgood and I did the best I could with
> > everyone's preferred times, and the best is that we could satisfy about
> > 90% of people when meeting at 1400 UTC on Wednesday, August 28th.
> >
> > This time, we can try to use Zoom:
> >
> > https://zoom.us/j/553650262
> >
> > However, I only have a free Zoom account, so we'll only get 40 minutes
> > to meet (and then we can just rejoin and continue if needed).
> >
> > Here was my idea for a basic agenda, but we don't have to stick to it.
> > :)
> >
> > - How is everyone doing? :) What is everyone up to?
> >
> > - mlpack development update
> > * I'll put together some slides about what's happened since the last
> > meeting, just like I did last time, and highlight developments and
> > changes
> >
> > - Release planning:
> > * what should go into the next release
> > * when should the next release happen
> > * should it be mlpack 3.2.0, 3.1.2, etc.?
> > * what are the states of open PRs, and what do we want to do with
> > them?
> >
> > - GSoC update/retrospective:
> > * we can talk about the status of each project and what we're hoping
> > for moving forward
> >
> > - Open mic topics (just some ideas):
> > * What should we change (if anything) about how we develop software
> > as a group?
> > * Is there anything we could improve? Is everyone having a good
> > time?
> > * Are there any big problems that we are overlooking that need to be
> > addressed, but aren't on Github as issues?
> > * Are there any big problems that we are overlooking that *are* on
> > Github as issues?
> >
> > Hope to see you there! And if you can't make it, we can send out some
> > notes of discussion afterwards so that nobody misses anything.
> >
> > Thanks!
> >
> > Ryan
> >
> > --
> > Ryan Curtin | "I like this game."
> > ryan at ratml.org | - Coach
> > _______________________________________________
> > mlpack mailing list
> > mlpack at lists.mlpack.org
> > http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
> >
>
> --
> Ryan Curtin | "Hungry."
> ryan at ratml.org | - Sphinx
> _______________________________________________
> mlpack mailing list
> mlpack at lists.mlpack.org
> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
>
--
Ryan Curtin |
ryan at ratml.org | "Death is the road to awe."
More information about the mlpack
mailing list