[Top][All Lists]

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

Re: [Gneuralnetwork] Welcome to Gneural Network

From: Jean Michel Sellier
Subject: Re: [Gneuralnetwork] Welcome to Gneural Network
Date: Wed, 16 Mar 2016 08:25:46 +0100

The Guile project sounds interesting but the thing is that my scripting language is almost done.. which is a good thing because we don't want to wait too long, right?


2016-03-16 7:49 GMT+01:00 Nala Ginrut <address@hidden>:
Hi Jean!

On Tue, 2016-03-15 at 17:03 +0100, Jean Michel Sellier wrote:
> Very soon, the following features will be released (I am in the process of
> rewriting some code that I already developed and tested):
> - a scripting language, so that the user can define and run a neural
> networks without even knowing how to program,

As a official GNU project, it's naturally to use GNU Guile as extension
scripting platform. The main language is Scheme, but a good news is that
the Lua frontend which I'm working on maybe released in this year. It
means that once you use GNU Guile API for extension, you have at least
Scheme and Lua as scripting languages. And the modules written in Scheme
or Lua could be applied interactively.
One effort, more results. Python? Maybe later. ;-)
To whom may concern, here's the project:

> - a repository will be created to make interactions easier,

Savannah is definitely the main choice, but Savannah supports many VCS
system(CVS/hg/git/svn...), I recommend we just use Git. It is necessary
to apply a repo on Savannah with manually review by other GNU folks.
Please don't hesitate to ask if you encounter any problem.

Besides, personally I prefer github as a mirror. For better maintain,
you may need a proper config in local git repo to push the commits to
github/savannah both, automatically.

I'm using both mailing-list and github issues to communicate with users.
But it depends on you to use mailing-list only.

> - recurrent networks, in particular LSTM neurons,
> - an automatic algorithm which select the best network topology given a
> specific problem,

I'm confusing here, do you mean an abstract mathematical network
topology? or detecting the fastest physical network node for connection

> - parallelization of the training process for multi-core CPUs and GPUs,

I'm not expertise on GPU, but if you want to use CUDA for parallel, it
seems impossible on Nouveau at present.
But someone is working on it:

Alternatively, we may choose OpenCL, but Nouveau seems not prepared yet:

But we still have some hope:

Anyway, we have two choices:
1. Using propriety GPU driver temporally, not good for GNU ethic, but
practical.  Waiting for the Nouveau.

2. Using CPU for it, and waiting for the Nouveau. The problem is the
time, Gneural-network folks may lose their patient.

I may get wrong, comments are welcome.

Best regards.

GNU Powered it
GPL Protected it
GOD Blessed it

HFG - NalaGinrut

Fingerprint 77F8 7D5B B425 E505 535E  C365 9182 7E28 EE78 E925


reply via email to

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