For historical reasons that I can't quite remember, we made it so that all
of the network types took exactly the same initial parameters, but then
allow setting additional properties before the network is actually
created. For example, we could have made it so that the constructor also
took layer names and sizes and actually made the layers. But we decided to
not allow that, and broke up initialization of the network from the
building of the network.

So, (as mentioned in the previous email) you can accomplish the same
affect with:

network = Network()
network.setSeed(38473)
# create network, etc

But, your suggestion to add the seed to the constructor makes a lot of
sense, is intuitive, should be backwards compatible, and shouldn't cause
any problems with networks that extend the Network class. I'll add this to
the CVS version, and test it out.

Thanks!

-Doug

On Fri, April 27, 2007 2:12 am, belinda thom said:
> Hi,
>
> Is there a cleaner way to set conx's seed so that I can run a
> repeatable training experiment?
>
> I ended up modifying conx's Network class's __init__, adding a seed flag
to it, e.g.
>
>      def __init__(self, name = 'Backprop Network', verbosity = 0,
> seed=None):
>          """
>          Constructor for the Network class. Takes optional name and
verbosity arguments.
>          """
>          if seed==None :
>              x = random.random() * 100000 + time.time()
>          else :
>              x = seed
>
> but would rather use a built-in method if there is one (I couldn't find
it). Otherwise, if such a thing doesn't exist, it would be worth making
this change so that, for teaching purposes, we can run a
> particular experiment in the classroom.
>
> Advice always appreciated.
>
> Thx,
>
> --b
>
>
> _______________________________________________
> Pyro-users mailing list
> [email protected]
> http://emergent.brynmawr.edu/mailman/listinfo/pyro-users
>


-- 
Douglas S. Blank
Associate Professor, Bryn Mawr College
http://cs.brynmawr.edu/~dblank/
Office: 610 526 6501



-- 
Douglas S. Blank
Associate Professor, Bryn Mawr College
http://cs.brynmawr.edu/~dblank/
Office: 610 526 6501

_______________________________________________
Pyro-users mailing list
[email protected]
http://emergent.brynmawr.edu/mailman/listinfo/pyro-users

Reply via email to