Anomaly score calculation separated to own function

[Zbysekz]

I've just moved anomaly score calculation in TemporalMemory::compute function to separate function and enriched py bindings by binding to that function.

I need this, to be able to calculate anomaly score if i am calling activateDendrites and activateCells separately as discussed briefly in #658

[breznak]

@Zbysekz code-wise the PR is good, a small change below
I'm somewhat reluctant to this change, as it mixes in Anomaly functionality into TM, which should have stayed separated,

• TM.anomaly is a convenience and only works (as stated in doc) if you use compute() only in your code
• for other scenarios you're on your own and have Anomaly, AnomalyLikelihood classes at your service (with full customization)

So the possible outcomes I see:

• rejecting the PR and using an instance of anomaly yourself (imho clearer design, unless too much extra code?)
• some compromise and making this function private in c++, and being more benevolent in py and exposing it there.

What do you think about this, conceptually? Or do you see a better way?

[breznak]

What I'll definitely like is the separation to a (private) method. The question is about exposing the functionality to outside world.

• one look is "if it can do that, allow us to use it"
• other "one class should do one thing well", and therefore is TM, Anomaly.

[Zbysekz]

Well i just did it this way because it was easier way.

I will modify this function according to your comments & delete pybind.

But you are right, that we shouldn't mix TM and Anomaly.
So how can i use Anomaly ? It seems that Anomaly py bindings is missing, so i should create one?

[breznak]

So how can i use Anomaly ? It seems that Anomaly py bindings is missing, so i should create one?

for c++ we have in src/htm/algorithms/

• Anomaly.hpp
• AnomalyLikelihood.hpp (note: implementation may be broken, see review and refactor AnomalyLikelihood code (both c++, python)  #469 )

in python:

• TM.anomaly (works OK for "RAW", maybe broken for other modes). But as you say, it's not for a flexible usecase.
• there is python-only (and working correct) py/htm/algorithms/anomaly_likelihood.py

So:

• it would definitely be nice to make bindings for Anomaly.hpp 👍
• maybe also do bindings for the c++ likelihood (duplicates the pure python file), and we'll have to compare/fix the c++ implementation

[breznak]

it would definitely be nice to make bindings for Anomaly.hpp +1

alternatively, only write the class in pure python, it should be quite easy implementation with numpy or SDR

[breznak]

So as discussed:

• keep the separate c++ function, but make it private
• implement Anomaly for python, either as bindings, or as a pure python code
  • and remove the bindings for TM related to anomaly in this PR
• optional: bindings for Likelihood (already pure python impl exists) + help verifying/fixing the c++ implementation

[breznak]

this one change and this looks good.
Will you want to address the py/bindings for anomaly as a part of this PR too?

[Zbysekz]

I will let you know, i think i will create anomaly calculation in pure python - at the end it is just simple calculation. I don't think that for one simple function is the effort of creating binding reasonable.

[breznak]

A few comments on the new Anomaly py implementation, please have a look below

[Zbysekz]

@breznak I commited this just because i am switching from laptop to desktop PC and i am now being interrupted often (my 2month old son :) ) so this is not yet to be merged, but i take your suggestions

[breznak]

I commited this just because i am switching from laptop to desktop

ok, seemed to me somehow half-aced.. ;) Enjoy the time with the baby!

[Zbysekz]

@breznak thanks. Maybe you can help me, suddenly when i compile htm.core i get error "Python.h" no such file or directory .. i know that this is related to missing python-dev (or python3-dev) but i have this installed... and previously it the compilation was successfull, do you have any advice?

[ 44%] Building CXX object bindings/py/cpp_src/CMakeFiles/sdr.dir/bindings/sdr/sdr_module.cpp.o In file included from /media/Data/Data/HTM/htm.core/bindings/py/cpp_src/bindings/sdr/sdr_module.cpp:24:0: /media/Data/Data/HTM/htm.core/bindings/py/cpp_src/bindings/suppress_register.hpp:35:10: fatal error: Python.h: No such file or directory #include <Python.h> ^~~~~~~~~~ compilation terminated.

[breznak]

i know that this is related to missing python-dev (or python3-dev) but i have this installed...

try removing the build/ folder and starting afresh? I'm also using venv to ensure the pythons dont mix,...

[Zbysekz]

ready to merge

[breznak]

This looks good, a few changes below, the biggest concern is

• does not work for multi-dimensional TM/SP,

[Zbysekz]

Ok, now should be like you suggested

[breznak]

This does look good now, thank you!

Please verify the 2 remaining concerns below and we're good to merge

[breznak]

Thanks for the improvements, let's merge this!
Thank you for this contribution, @Zbysekz 👍