Thanks for the pointers, Matt! I watched your tutorial video and follower your steps myself. One weird thing was that when I did the swarm over the sine data set, no model_params.py was generated. I checked out the exact example code you posted on your personal GitHub, so we will see if that happens with your exact code. Also of note is that the description.py did not seem to work with Cerebro.
Thank you for pointing to the skeleton doc about model params. That is a good start, and in line with what I was looking for. I wasn't able to find that on my own, so hopefully the reorg will help. The other doc I feel is missing is one that says "here are the major pieces of py code and how they fit together." Not necessarily of interest to the casual user, but could be immensely helpful to prospective contributors. On Wednesday, March 26, 2014, Matthew Taylor <[email protected]> wrote: > Most of the Numenta team is off at an AWS event today, so I'll try to > help as I can. We are lacking documentation on model parameters, which > is bad because they are quite important. > > The swarm process seems magical, but all it does is run a PSO [1] > algorithm. Hopefully, you've seen Ron's video on Swarming [2]. So it > just takes your input data and creates a bunch of models over and > over, throwing out the ones that don't perform well and replacing them > with new models with different params. Eventually finding the set of > parameters that perform best on your data. More on our wiki [3] [4]. > > We do have some docs on model params on our wiki as well [5], but it > needs to be better exposed and cleaned up. (I'm working on wiki reorg > right now [6]). > > [1] en.wikipedia.org/wiki/Particle_swarm_optimization > [2] Swarming in NuPIC <http://www.youtube.com/watch?v=xYPKjKQ4YZ0> > [3] https://github.com/numenta/nupic/wiki/Running-Swarms > [4] https://github.com/numenta/nupic/wiki/Swarming-Algorithm > [5] https://github.com/numenta/nupic/wiki/Models#version-1 > [6] https://github.com/numenta/nupic/issues/747 > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > On Wed, Mar 26, 2014 at 2:57 PM, Julie Pitt <[email protected]<javascript:;>> > wrote: > > Yes. I read On Intelligence and the Whitepaper. Now I actually want to > take > > my understanding to the next level. What I've worked out so far is that > > there are a few options, for running a model: > > > > 1. Use the OPF. You're essentially a (python) client to the OPF which in > > turn is a client to the CLA. > > 2. Use python (or other) language bindings directly to interface with the > > CLA > > > > I'm currently looking at #1. There are several ways I've seen to get into > > using the OPF: > > > > a. Given a dataset, swarm it to create a model. Somehow swarming > *magically* > > figures out what translation and encoding is needed, whether you need to > use > > the SP or TP or both, as well as what parameters to send them. > > b. create a model by hand > > > > I can do the swarm but would really like to understand the various > > parameters and config in description.py as well as other elements I > might be > > missing. Right now I'm blindly fumbling around until it starts to make > > sense. I'd love to do better than that if others have advice. > > > > > > > > On Wed, Mar 26, 2014 at 2:46 PM, Freeman 77 > > <[email protected]<javascript:;>> > wrote: > >> > >> Did you checked the available documents in the wiki about the CLA and > the > >> book On Intelligence? I think this must be the first step you need to > make, > >> at least that's what I'm doing. > >> > >> Greetings. > >> > >> El 26-03-2014, a las 17:50, "Julie Pitt" <[email protected]<javascript:;>> > escribió: > >> > >> Sorry if this post appears twice. I used the wrong "from" address the > >> first time. > >> > >>> I'm just getting into NuPIC. I have built it on my machine and I'm > >>> running Cerebro. I've tried loading up a couple of models (so far only > >>> hotgym works) and run using Cerebro. I would like to start tinkering by > >>> tweaking or creating my own models. Aside from just reading > examples/code > >>> and making inferences, is there a good place to go to get a > description of > >>> the config and what the params are? > >>> > >>> I have stumbled across ExpGenerator.py which seems to generate > >>> experiments from some input file, but so far I haven't been able to > >>> determine what should be in that input file. > >>> > >>> Thanks! > >> > >> > >> > >> _______________________________________________ > >> nupic mailing list > >> [email protected] <javascript:;> > >> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > >> > >> > >> _______________________________________________ > >> nupic mailing list > >> [email protected] <javascript:;> > >> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > >> > > > > > > _______________________________________________ > > nupic mailing list > > [email protected] <javascript:;> > > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > > > _______________________________________________ > nupic mailing list > [email protected] <javascript:;> > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
