Belinda,

You can use network.reset() to start the pseudo-random stream of numbers
back to the beginning with the original seed, and then re-assign the
weights again. network.initialize() will just re-assign the weights in the
current position in the pseudo-random stream.

So, you can rerun experiments over and over again with:

network.setSeed(38736)
network.train()
network.reset()
network.train()
...

Or, if you wish to run a bunch of repeatable experiments in which no two
are the same:

network.setSeed(38736)
network.train()
network.initialize()
network.train()
network.initialize()
network.train()
...

BTW, conx prints out the seed number regardless of verbosity for a related
reason. When I was in grad school, Jordan Pollack told us a story about
this wonderful neural network that produced this beautiful "butterfly
pattern" in weight space (you can probably can find the image by reading
some of his papers from the early '90s). Well, he didn't save the initial
seed of this network, and never (as far as I know) found another network
like it, although he did spend some time trying. So, I always know what
the initial seed is, just in case I want to go back and run that again.

network.resetFlags() should be a hidden/private method; it is used to make
sure that inputs and activations are set/computed appropriately (for
example, it checks to make sure that you set all of the inputs once and
only once).

-Doug


On Fri, April 27, 2007 2:52 am, belinda thom said:
> Hi,
>
> I've been writing my own train function for a variety of reasons and
> have recently wanted to do several restarts so that I can see which
> gives the best result.
>
> What I can't figure out how to do is to set the weights before each
> attempt to train so that they are re-randomized. I had hoped the
> network.resetFlags() that conx's Network class's train method used
> would do the trick, but it doesn't re-randomize the initial weights.
>
> Any suggestions?
>
> Thanks,
>
> --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

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

Reply via email to