Re: Is theano worth saving?
On Wed, 2017-02-08 at 10:58 +0100, Joost van Baal-Ilić wrote:
> Hi,
>
> On Wed, Feb 08, 2017 at 09:47:08AM +0000, Ghislain Vaillant wrote:
> > On Wed, 2017-02-08 at 10:17 +0100, Daniel Stender wrote:
> > > On 08.02.2017 09:13, Rebecca N. Palmer wrote:
> > > > I have patches for #848764 and #831540; I haven't yet found the cause of
> > > > #831541, but as it only affects big-endian systems, partial removal is an
> > > > option.
> > > >
> > > > However, one of these bugs is a point where upstream plan to make an
> > > > incompatible change (though my fix doesn't), and upstream aren't sure
> > > > whether having Theano in a long-release-cycle distribution such as Debian
> > > > is a good idea: https://github.com/Theano/Theano/issues/5494
> >
> > On first read, it sounds to me that upstream is failing to understand
> > that having theano in the archive does not replace or substitute
> > traditional development workflow via pip if users want to.
>
> And/or maybe upstream feels anybody using an "obsolete" version of theano is
> doing something wrong. And/or maybe upstream would feel some kind of
> obligation/duty to support users of "obsolete" versions of Debian-shipped
> theano on Debian systems, if Debian were to ship with those. See also the
> xscreensaver debacle, some months ago. I'm just guessing here, extrapolating
> from other upstreams.
xscreensaver is an application, so upstream's anger was justified since
users had no other (simple) alternative to using the outdated packaged
version.
theano is a library, and libraries provided in Debian should be
packaged as redepends for other frameworks / applications . That's why
you are invited to provide a justification in your ITPs as to why
packaging the library for Debian is necessary.
Unlike with xscreensaver, users do have an alternative for using the
latest theano: fire a virtualenv and use the latest theano there via
`pip install`.
Same comment for numpy, scipy and the rest of the scientific Python
ecosystem. Anyone using for instance jessie + the system numpy for
serious development is clearly missing out, and should get to know
modern development tools (i.e. venv, pip, anaconda...).
> > Our packages are here to support future applications / frameworks,
> > which may require an appropriately packaged theano. Considering the
> > success of deep-learning these days, this prospect is quite likely.
> >
> > > > (Note that their suggested alternative of using pip will involve manually
> > > > installing libblas-dev, as pip only knows about Python dependencies.)
>
> Anyway, it would be good to make clear what Debian expects of upstream, and how
> shipping theano with Debian might even be _beneficial_ for upstream.
Indeed, in the form of dynamic CI (with autopkgtest) and testing on a
variety of architectures they might not have access to otherwise.
Whether they care or not is a different story.
I believe the easiness of `apt install` which Daniel brought in is no
longer so much of a *strong* argument, now that pip + wheels has become
quite mature. That's my personal opinion though.
Ghis
Reply to: