John, > On Jul 16, 2015, at 10:29 AM, John Blackburn <[email protected]> > wrote: > > Dear All > > I've now prepared a basic temporal memory example based on the > "hello_tp.py" test for the Temporal Pooler. it does the same thing, > learns the ABCDE sequence and outputs in a similar format. I think > this example might be useful for beginners so feel free to add it to > the repo if you like.
Great, thanks! > > The TM is able to learn the sequence perfectly, like TP did. However, > the choice of which cell to activate in each column seems to be random > now, whereas with TP it was always the bottom cell in each column. > Also, the TM is "surprised" by the first input with all cells active > in each activated column. That behavior is expected and correct. > > In the next step I'll try to make it lean a simple sinusoid sequence. > > John. > > On Wed, Jul 15, 2015 at 9:40 AM, John Blackburn > <[email protected]> wrote: >> Thanks very much, Matthew. This will help many people. >> >> John >> >> On Tue, Jul 14, 2015 at 5:35 PM, Matthew Taylor <[email protected]> wrote: >>> Fixed! http://numenta.org/docs/nupic/ >>> --------- >>> Matt Taylor >>> OS Community Flag-Bearer >>> Numenta >>> >>> >>> On Tue, Jul 14, 2015 at 9:23 AM, John Blackburn >>> <[email protected]> wrote: >>>> Looks like it might be time to run doxygen again! Last run in May 19. >>>> >>>> John. >>>> >>>> On Tue, Jul 14, 2015 at 5:14 PM, cogmission (David Ray) >>>> <[email protected]> wrote: >>>>> Hey John, >>>>> >>>>> Nice self-sufficient researching! I like that in ya' !!! >>>>> >>>>> Anyway, yes that last (stripNeverLearned) parameter was recently removed >>>>> last month. The file I gave you is older than that... >>>>> >>>>> Remember, NuPIC is ever evolving, and it is still technically >>>>> "pre-release"! >>>>> >>>>> ;-) >>>>> >>>>> >>>>> >>>>> On Tue, Jul 14, 2015 at 11:09 AM, John Blackburn >>>>> <[email protected]> wrote: >>>>>> >>>>>> Thanks, Ralf, >>>>>> >>>>>> Actually that reminds me, David Ray kindly sent me the QuickTest.py >>>>>> example in Python so I just tried that. However, I ran into another >>>>>> problem: there seems to be some confusion about how many parameters >>>>>> sp.compute() takes (Spatial Pooler). In QuickTest.py the code reads >>>>>> >>>>>> sp.compute(encoding, True, output, False) >>>>>> >>>>>> However, on Github sp.compute takes only 3 parameters (apart from self): >>>>>> >>>>>> >>>>>> https://github.com/numenta/nupic/blob/master/nupic/research/spatial_pooler.py#L658 >>>>>> >>>>>> So this causes a crash. I notice on the API docs the 4th parameter is >>>>>> indeed mentioned: >>>>>> >>>>>> >>>>>> http://numenta.org/docs/nupic/classnupic_1_1research_1_1spatial__pooler_1_1_spatial_pooler.html#aaa2084b96999fb1734fd2f330bfa01a6 >>>>>> >>>>>> So I guess the 4th arg was recently removed. Pretty confusing! >>>>>> >>>>>> Can anyone shed light on this mystery? >>>>>> >>>>>> John. >>>>>> >>>>>> On Tue, Jul 14, 2015 at 12:25 PM, Ralf Seliger <[email protected]> wrote: >>>>>>> Hey John, >>>>>>> >>>>>>> why don't you try the QuickTest example in htm.java >>>>>>> (https://github.com/numenta/htm.java) or htm.JavaScript >>>>>>> (https://github.com/nupic-community/htm.JavaScript)? It involves the new >>>>>>> temporal memory, and stepping through the code with a debugger you can >>>>>>> easily study the inner workings of th algorithm. >>>>>>> >>>>>>> Regards, RS >>>>>>> >>>>>>> >>>>>>> Am 14.07.2015 um 11:39 schrieb John Blackburn: >>>>>>>> >>>>>>>> Thanks, Chetan, >>>>>>>> >>>>>>>> Any tutorials, examples of how to use temporal_memory.py? The nice >>>>>>>> thing about old TP is it has an example: hello_tp.py. >>>>>>>> >>>>>>>> John. >>>>>>>> >>>>>>>> On Mon, Jul 13, 2015 at 7:55 PM, Chetan Surpur <[email protected]> >>>>>>>> wrote: >>>>>>>>> >>>>>>>>> Hi John, >>>>>>>>> >>>>>>>>> The TP is now called "Temporal Memory", and there's a new >>>>>>>>> implementation >>>>>>>>> of >>>>>>>>> it in NuPIC [1]. Please use this latest version instead, and let us >>>>>>>>> know >>>>>>>>> if >>>>>>>>> you still find issues with the results. >>>>>>>>> >>>>>>>>> [1] >>>>>>>>> >>>>>>>>> >>>>>>>>> https://github.com/numenta/nupic/blob/master/nupic/research/temporal_memory.py >>>>>>>>> >>>>>>>>> Thanks, >>>>>>>>> Chetan >>>>>>>>> >>>>>>>>> On Jul 13, 2015, at 4:44 AM, John Blackburn >>>>>>>>> <[email protected]> >>>>>>>>> wrote: >>>>>>>>> >>>>>>>>> Dear All >>>>>>>>> >>>>>>>>> I'm trying to use the temporal pooler (TP) directly as I want to get >>>>>>>>> into the details of how Nupic works (rather than high level OPF etc) >>>>>>>>> >>>>>>>>> Having trained the TP I used this code to get some predictions: >>>>>>>>> >>>>>>>>> for j in range(10): >>>>>>>>> x=2*math.pi/100*j >>>>>>>>> y=math.sin(x) >>>>>>>>> >>>>>>>>> print "Time step:",j >>>>>>>>> >>>>>>>>> for k in range(nIntervals): >>>>>>>>> if y>=ybot[k] and y<ytop[k]: >>>>>>>>> print "input=",x,y,k,rep[k,:] >>>>>>>>> >>>>>>>>> tp.compute(rep[k,:],enableLearn=False,computeInfOutput=True) >>>>>>>>> tp.printStates(printPrevious = False, printLearnState = >>>>>>>>> False) >>>>>>>>> break >>>>>>>>> >>>>>>>>> >>>>>>>>> Here is the result I got: >>>>>>>>> >>>>>>>>> Time step: 0 >>>>>>>>> input= 0.0 0.0 9 [0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0] >>>>>>>>> >>>>>>>>> Inference Active state >>>>>>>>> 0000000001 0000000000 >>>>>>>>> 0000000000 0000000000 >>>>>>>>> Inference Predicted state >>>>>>>>> 0000000000 0000000000 >>>>>>>>> 0000000001 0000000000 >>>>>>>>> Time step: 1 >>>>>>>>> input= 0.0628318530718 0.0627905195293 10 [0 0 0 0 0 0 0 0 0 0 1 0 0 0 >>>>>>>>> 0 0 0 0 0 0] >>>>>>>>> >>>>>>>>> Inference Active state >>>>>>>>> 0000000000 1000000000 >>>>>>>>> 0000000000 0000000000 >>>>>>>>> Inference Predicted state >>>>>>>>> 0000000000 0000000000 >>>>>>>>> 0000000001 0000000000 >>>>>>>>> Time step: 2 >>>>>>>>> input= 0.125663706144 0.125333233564 11 [0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 >>>>>>>>> 0 0 0 0 0] >>>>>>>>> >>>>>>>>> Inference Active state >>>>>>>>> 0000000000 0100000000 >>>>>>>>> 0000000000 0100000000 >>>>>>>>> Inference Predicted state >>>>>>>>> 0000000000 0000000000 >>>>>>>>> 0000000000 1110000000 >>>>>>>>> Time step: 3 >>>>>>>>> input= 0.188495559215 0.187381314586 11 [0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 >>>>>>>>> 0 0 0 0 0] >>>>>>>>> >>>>>>>>> Inference Active state >>>>>>>>> 0000000000 0000000000 >>>>>>>>> 0000000000 0100000000 >>>>>>>>> Inference Predicted state >>>>>>>>> 0000000000 0000000000 >>>>>>>>> 0000000000 1110000000 >>>>>>>>> >>>>>>>>> You can see that in time step 3, one cell (12th column) is shown as >>>>>>>>> being both in the active and predictive state, which I though was >>>>>>>>> impossible. (its inference active state is 1 and its inference >>>>>>>>> predicated state is 1) >>>>>>>>> >>>>>>>>> Also if you look at time step 0, only 1 cell is in the predictive >>>>>>>>> state. However, the input that comes in at time step 1 activates the >>>>>>>>> colum to the right of this cell (the 11th slot is "1") so I would >>>>>>>>> expect the 11th column to have both cells active, the "unexpected >>>>>>>>> input state" but this does not happen. >>>>>>>>> >>>>>>>>> Can anyone explain this? >>>>>>>>> >>>>>>>>> John. >>>>>>>>> >>>>>>>>> >>>>>>> >>>>>>> >>>>>> >>>>> >>>>> >>>>> >>>>> -- >>>>> With kind regards, >>>>> >>>>> David Ray >>>>> Java Solutions Architect >>>>> >>>>> Cortical.io >>>>> Sponsor of: HTM.java >>>>> >>>>> [email protected] >>>>> http://cortical.io >>>> >>> > <hello_tm.py>
