guix-patches
[Top][All Lists]
Advanced

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

[bug#36071] [PATCH] gnu: Add leela-zero.


From: Ricardo Wurmus
Subject: [bug#36071] [PATCH] gnu: Add leela-zero.
Date: Tue, 11 Jun 2019 22:55:45 +0200
User-agent: mu4e 1.2.0; emacs 26.2

Ludovic Courtès <address@hidden> writes:

>> Ludovic Courtès <address@hidden> a écrit :
>>
>>> Does the game run at all if we don’t download additional network
>>> weight data?
>>>
>>> There’s a debate as to whether trained neural network parameters can
>>> be considered “source”: <https://lwn.net/Articles/760142/>.
>>
>> The game needs neural network weights to run. The user must download
>> a file with the weights (usually from
>> <https://zero.sjeng.org/best-network>, which is updated daily I think)
>> and put it in '$HOME/.local/share/leela-zero/'.
>>
>> Do you think some weigths (e.g. taken from
>> <https://leela.online-go.com/networks>, for reproducibility) should be
>> embedded in the package?
>
> My point is about whether these trained neural network data are
> something that we could distribute per the FSDG.
>
> The LWN article above mentions ‘leela-zero’.  In Debian (which has a
> different policy), ‘leela-zero’ has apparently been accepted, but
> without the data, and with a link to said data (as in your original
> patch):
>
>   https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=903634
>
> Perhaps we could do the same, but I’d like to hear what others think.

We need to ensure that the software necessary to train the networks is
included.  Is this the case?

Distributing the training data is not an option, as I see it, because
the files are very large.

I don’t know about the networks that are offered for download at
https://leela.online-go.com/networks/.   These are files of very
different sizes, and it is neither clear what license applies nor how
they were arrived at.  They are zipped up magic numbers.

Arguably a similar statement could be made about impulse response files
for audio filters (such as convolution reverbs).  While we offer
convolution filters (such as the “ir” package), we don’t include data
files.  Had I not made the link to network weights for ML models I would
have considered impulse response files no different from e.g. game
assets.  Now I’m not so sure any more.

I don’t really know what it really means for data to be
“non-functional”.

Back to this patch: I think it’s fine to accept it as long as the
software necessary for training is included.

--
Ricardo






reply via email to

[Prev in Thread] Current Thread [Next in Thread]