Re: recording activation of hidden units
(Që… ð@ÄØð@ÄÔ)
Tue, 12 Aug 1997 11:57:51 +0200 (MET DST)
> > I'm setting up a three-layered feedforward net with SNNS. I want to record the
> > hidden unit activations and have their activation listed in a separate data file
> > to use as distributed input on the same net a second time round.
> >
> > Adam Aron
>
> The only method I can think of is to train the whole network then to
> use the relevant weights and build the reduced net having as output
> layer the hidden one you are interested in.
>
I am afraid, that the only method is to rebuild the snns source code. I have done it
already, both to write down the hidden layer activations, as well as to add new learning
methods. I think it is better to spend some time ONCE to print relevant parts of snns
source, then read it and modify it. A method of first learning, then rebuilding the net as
to have the hidden layer as output one is cumbersome - you have to employ the graphics net
editor, it takes a long time and is prone to errors. And, last but not least, you have to
do it every time. It really is better to change the code.
igor t. podolak
---------------------------------
igor t. podolak
Institute of Computer Science
Jagiellonian University
31-127 Krakow
e-mail: uipodola@jetta.if.uj.edu.pl
igor@neuron.ii.uj.edu.pl