gneuralnetwork
[Top][All Lists]
Advanced

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

Re: [Gneuralnetwork] Adding a build target?


From: Ray Dillinger
Subject: Re: [Gneuralnetwork] Adding a build target?
Date: Thu, 20 Oct 2016 00:02:37 -0700
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Icedove/45.4.0


On 10/18/2016 09:38 PM, Ray Dillinger wrote:

> We need to name it.  Right now I'm torn between gnnet, nnet, and zorgo,
> but I'm open to ideas if anyone else has one.

Well, I didn't hear anything immediately, so I went ahead and checked in
changes.  The new executable is 'nnet.'  The new file (in src/) is nnet.c.

Please beat on the new parser/writer combo.  nnet is supposed to be
fully transparent - ie, when you do something wrong it gives you a
relevant, explicit error message that gives you information about what's
going on and how to do it right.  Test this!  Start knowing nothing
about the language it accepts and see if you can build a valid script
based on nothing but error messages from your attempts.

The functionality is pretty limited right now; it reads, parses, and
error checks a network file, building an internal network representation
- then turns around and writes that network representation out to an
output file.

We're starting with writeback capability; you can write networks out and
read them back in, in the same configuration language you use to define
them.  Which means we can save, edit the network as a text file, reload,
and continue training or deploy with changes. This will be really useful
once the network file includes coming features such as training plans
and data sources.  At the very least it will mean training needn't start
over every time parameters change.

The output file if read by the parser produces a network having the same
topology, weights, and activation sequence.  It is completely
reconstructed from the internal representation without reference to the
original input file or its structure, so if the network structure
changes during training (due to genetic algm, simulated annealing, or
whatever arbitrary transformation), the output will correctly record the
changes.

                                Bear

Attachment: signature.asc
Description: OpenPGP digital signature


reply via email to

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