[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: [covid-19] Reviving tensorflow packaging effort (Was: Missing dependancies for streamlit)





On Sat, Apr 25, 2020, 12:36 Mo Zhou <lumin@debian.org> wrote:
Hi Tille,

Is there any COVID-19 package using pytorch blocked due to its absense?

A good news is that I've managed to strip the whole third_party/
directory of src:pytorch, and started to forward my patches to upstream[1].
When all my modifications entered the upstream repo, I'll be quite
confident that our src:pytorch package can enter the archive without
any (annoying) embedded sources [2].

What I'm doing now is to wait for the upstream to merge my commits, and
for the ftp-masters to accept my NEW dependency packages.

In that sense, I'd like to take the COVID-19 shortcut to pass NEW
quickly, if any COVID-19 related package needs pytorch.

--- According to my status page
https://qa.debian.org/developer.php?login=lumin
these are the NEW dependency packages:

fp16
fxdiv
gloo
onnx
psimd
pthreadpool

Go ahead and add them to 
https://salsa.debian.org/med-team/community/2020-covid19-hackathon/-/wikis/NEW-Requests 


I'll start to re-debianize src:pytorch from scratch when all of my
commits had been upstreamed.

Any reason to not add the removed files/folders to the Files-Excluded stanza of debian/copyright and repack the source while you wait?


--- all of my on-going work are publically available:
https://salsa.debian.org/deeplearning-team
https://github.com/cdluminate/pytorch (messy, not rebased yet)
The links to my PRs can be found on [1]

[1] https://github.com/pytorch/pytorch/issues/14699
[2] Finally, we will have a modern deep learning framework in the
    archive. It's better than having nothing even if I'm working on the
    cpu-only (free) version.

On Wed, Apr 08, 2020 at 10:06:12AM +0200, Andreas Tille wrote:
> Hi Mo,
>
> On Wed, Apr 08, 2020 at 02:46:06AM +0000, Mo Zhou wrote:
> > On Tue, Apr 07, 2020 at 11:49:07AM +0200, Andreas Tille wrote:
> > > On Tue, Apr 07, 2020 at 08:56:45AM +0100, Rebecca N. Palmer wrote:
> > > > tensorflow 1.10 was packaged in experimental, but with reduced performance,
> > > > and was removed because this was considered not worth it:
> > > > https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=935769
> > >
> > > Ah I simply forgot this.  Thanks for refreshing my mind.
> >
> > After that I tried to refresh the packaging and uploaded tensorflow 2.0
> > to the NEW queue. The ftp-master complained about the embedded snapshot
> > version of Eigen3. Ftp-masters are not convinced even if I said
> > tensorflow FTBFS against the version shipped in our archive, and it
> > could waste lots of my time and energy to patch the related code.
>
> I can understand your feeling.  I've re-read the discussion[1] about
> including eigen3 into the tensorflow source.  The most promising
> statement was given in

>    https://alioth-lists.debian.net/pipermail/debian-science-maintainers/2020-March/079169.html
>
>    Sean Whitton spwhitton at spwhitton.name
>    Tue Mar 3 04:04:52 GMT 2020
>
>    Rejected per your request, but from my point of view this discussion is
>    not over -- Policy 4.13 says that packages *should* not use convenience
>    copies of code, not that they *must* not.
>
>    Thank you for all your work on uploading useful ML packages.
>
> At first: Thanks also from me!
>
> >From my point of view that is not a lost case so we should really try
> again.
>
> > I was so angry at that time so I deleted my name from the maintainers
> > and wrote the git message "I'm dropping this burden.".
>
> Well, sometimes personal feelings are dominating our actions.  I hope we
> could form some real team around this to spread the technical as well as
> the organisational burden.

> > So another kind notice for whoever is willing to take over tensorflow:
> > also be prepared to fuss with ftp-master.
>
> I need to admit that I'm absolutely happy about ftpmaster.  They are
> currently *extremely* supportive to our COVID-19 hackathon and > 30
> packages made it from upload to unstable in less than 24 hours!
>
> Since I consider it a "promising time" to reach a lot for deep learning
> tools in Debian which are frequently used in those tools to hunt down
> COVID-19 I'd like to call for help here for trying again to get
> tensorflow in - this time even with the Python3 module.
>
> Lumin has written an own build system for the C++ library since he has
> found out that the upstream build system is not usuable for Debian.
> Regarding the Python3 module he said that its not simply a matter of
> adapting his build system but "significantly extend" it since the python
> building process is much more complicated than the process for C++.
>
> Any takers for this task?
>
> Kind regards
>
>        Andreas.
>
>
> [1] https://alioth-lists.debian.net/pipermail/debian-science-maintainers/2020-March/079054.html
>
> --
> http://fam-tille.de
>


Reply to: