Hey Matt, I’m using the modelFactory at the moment and creating as many separate models as I need, each assigned to a single shape/gesture. Each model is trained by feeding in the sequences of a single class and resetting using model[i].resetSequenceStates() after each seq.
Testing is done using another partition of the data by feeding in test sequences with learning turned off and recording the resulting anomaly score at the end of the sequence. Used a small swarm using all the data and initialized all models using the same model parameters. I’ll have more formal results to share in terms of accuracy tomorrow. Thanks! Nicholas > On Dec 11, 2014, at 12:54 AM, Matthew Taylor <[email protected]> wrote: > > Nicholas, > > If I understand correctly, that else statement occurs when no > predictions are being made. How are you setting up your multiple HTM > regions. You must be using the Network API directly, right? > > Also, if you want to change the code, you'll need to re-build or use > "python setup.py develop". See > https://github.com/numenta/nupic#developer-instructions for more > details. > > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > On Wed, Dec 10, 2014 at 11:15 AM, Nicholas Mitri <[email protected]> wrote: >> Hey all, >> >> I keep having to work around this piece of code in the anomaly file : >> >> elif len(prevPredictedColumns) > 0: >> # There were predicted columns but none active. >> score = 1.0 >> else: >> # There were no predicted or active columns. >> score = 0.0 >> >> I have a setup of multiple HTM regions, each trained on sequences associated >> with a shape (sequences of stroke directions). When I run a test sequence >> through all of them and record the anomaly scores to find the minimum and >> therefore which shape is most likely being observed, I get 0 scores because >> of the else statement. I tried setting it to -1 but the files are all read >> only and no amount of chmod 777 is fixing it. >> Any other way to detect when anomaly is forced to zero? >> >> Thanks, >> Nicholas >
