Autolearning problem

Engineering NuPIC
1

Hello. There are some questions about learning model in hot gym prediction tutorial. Has enableLearning been turned on in this example or not? If not how can I turn enableLearning on? Is it enough to use something like model.enableLearning()?

2

If you create an OPF model, learning is enabled in both the SP and the TM by default. You can turn them both on and off with:

1 Like
3

And no model.finishLearning() needed right @rhyolight ?

Iā€™m turning off learning after a certain set point and leaving it off, so donā€™t want miss any way to optimize inference from there.

1 Like
4

Right, only call finishLearning() if you know you wonā€™t need to learn ever again as an inference optimization.

5

Ok great! So Iā€™m trying to use model.finishLearning right after model.disableLearning() and getting this error:

Traceback (most recent call last):
  File "<input>", line 623, in <module>
  File "<input>", line 465, in do_stream
  File "/Users/samheiserman/Desktop/Dppm.app/Contents/Resources/DPPM_app/myenv_folder/lib/python2.7/site-packages/nupic/frameworks/opf/htm_prediction_model.py", line 306, in finishLearning
    self._getSPRegion().executeCommand(['finishLearning'])
  File "/Users/samheiserman/Desktop/Dppm.app/Contents/Resources/DPPM_app/myenv_folder/lib/python2.7/site-packages/nupic/engine/__init__.py", line 467, in executeCommand
    return self._region.executeCommand(args)
  File "/Users/samheiserman/Desktop/Dppm.app/Contents/Resources/DPPM_app/myenv_folder/lib/python2.7/site-packages/nupic/bindings/engine_internal.py", line 1555, in executeCommand
    return _engine_internal.Region_executeCommand(self, *args, **kwargs)
  File "/Users/samheiserman/Desktop/Dppm.app/Contents/Resources/DPPM_app/myenv_folder/lib/python2.7/site-packages/nupic/bindings/regions/PyRegion.py", line 402, in executeMethod
    raise Exception('Missing command method: ' + methodName)
Exception: Missing command method: finishLearning

Hereā€™s the line itā€™s happening. Iā€™m streaming in multivariate files and maintaining separate models for each field, so the self.fields_modelsdicts[field][ā€˜modelā€™] 's are ModelFactory.model objects. This works with disableLearning() on its own, but not once I add finishLearning().

self.fields_modelsdicts[field]['model'].disableLearning()
self.fields_modelsdicts[field]['model'].finishLearning()

This performance gain could add up to a big difference as the system scales to many fields (why Iā€™m bugging you on this).

Thanks again @rhyolight!

6

@sheiser1 Is the model instance youā€™re using a TemporalMemory or BacktrackingTM?

1 Like
7

BacktrackingTM, the one that comes with getScalarMetricWithTimeOfDayAnomalyParams

8

My suspicion is that youā€™re using CPP version of the algorithm, and it doesnā€™t have a finishLearning function, which probably means it doesnā€™t need to be optimized in CPP (hopefully). Let me know if you notice a speed different, but I would just not call that function.

1 Like
9

Yep I am

tmImplementation="cpp"

Ok, so I could try it with ā€œtm_cppā€ and .finishLearning(), and see if thereā€™s a speed difference from the current setup.

Thanks again :smile:

10

Thereā€™s cpp and tm_cpp, which are unfortunately named, especially since I canā€™t remember which one is the TemporalMemory vs BacktrackingTM.

1 Like
11

The cpp is default in getScalarMetricsWithTimeofDayAnomalyParams(), I suspect its the BacktrackingTM

1 Like
12

Be aware that tm_cpp will not work as cpp well on anomaly detection problems.

1 Like
13

Do you know why this is or where I could go to find out?

14

Yes, it is because we added some non-biological hacks into the BacktrackingTM so it could better identify sequences. This helped with anomaly detection, but it did not move us towards a better understanding of the cortex. The TemporalMemory algorithm is not as performant at anomaly detection as the older non-biological one because it has not been ā€œhand-tunedā€ to any one task. (The whole point of our work being that we donā€™t want to hand-tune anything)

2 Likes