pyetaler

A high performance implementation of Numenta's HTM algorithms


Keywords
HTM, Hierarchical, Temporal, Memory, Numenta, AI, SDRsparse, distributed, representation, bioinspired
License
BSD-3-Clause
Install
pip install pyetaler==0.0.6

Documentation

PyEtaler

This is the offical Python binding for Etaler. PyEtaler generates Python binging using cppyy and adds additional feature on top of the automatically generated bindings.

Note: As of now, installing cppyy (thus PyEtaler) will cause ROOT to fail to load due to dependency clash.

Installation

Note: You must have Etaler and cppyy installed globally before building the binding. Note: Since the binding is generated to load the actual Etaler installation. You'll need to re-compile the binding everytime Etaler is updated.

If you are building from source (building via directly interacting with the generator).

pip install cppyy # must installed globally
python3 genbinding.py
cp *.so etaler/
cp *.pcm etaler/
# Then copy the resulting files into your package directory

Locally build via PIP

pip3 install .

Alternativelly you can install it directly via PyPI.

pip install pyetaler

Usage

After installation, you can use Etaler from python. The API is exactly like it is in C++.

>>> from etaler import et
>>> et.ones([2, 2])
{{ 1, 1}, 
 { 1, 1}}
>>> sp = et.SpatialPooler([128], [32])
>>> x = et.encoder.scalar(0.1, 0, 1, 128, 12)
>>> sp.compute(x)
{ 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0}

Caveats

PyEtaler tries it's best to make the Python API similar to the C++ API (so ou can use the C++ documents as the Python document). There are some caveats. A few changes are made in an effort to make the Python API Pythonic. The changes are:

Tensor.toHost and Tensor.item are not templates

Since Python is a dynamically typed language. There's really no point have functions maintaining a template. Your code in C++

auto t = et::ones({4,4});
auto vec = t.toHost<int>()

becomes

t = et.ones([4,4])
vec = t.toHost()

No more brace arround tensor indices

It is quite annoning having to have extra braces when indexing. So we removed them in Python!

auto t = et::ones({4,4});
auto q = t[{2, 2}];

becomes

t = et.ones([4,4])
q = t[2, 2]

Logical operators doesn't work

Due to how Python works. It is impossible to provide the logical operators in the wrapper.

auto r = t && q;

Should have become:

r = t and q

But instead of mapping operator &&. Python will try casting et.Tensor into a bool then performing the and operation. So instead of using and. Please use the logical_and function.

r = et.logical_and(t, q)

Hacking PyEtaler

In case that you need to use C++ STL - maybe because the wrapper is doing something stupid. You can access the STL using etaler.std.

For example

>>> from etaler import std
>>> std.vector[int](10)
<cppyy.gbl.std.vector<int> object at 0x1ef112c0>