GridCellsEncoder: simplify encode()

[breznak]

Currently, the encode expects a SDR parameter encode([x, y], sdr_gc) grid cells.

Used as:

SDR gc_sdr
Metrics(gc_sdr)
GridCellsEncoder gce
for d in data:
  gce.encode(d, gc_sdr)

I'm proposing GCE creating the SDR itself, and exposing it as a member GCE.grid_cells, so

GridCellsEncoder gce
Metrics(gce.grid_cells)
for ...:
  gce.encode([x,y])

Also consider if encode shouldn't take only x,y integer params encode(x, y),
as the 2D point/list [x,y] is the only input the encoder accepts and we have to check for it. This might be clearer.

Follow up to #353

[ctrl-z-9000-times]

Also consider if encode shouldn't take only x,y integer params encode(x, y),
as the 2D point/list [x,y] is the only input the encoder accepts and we have to check for it. This might be clearer.

I agree that it would be nice if this accepted coordinates as two arguments instead of a pair. I don't know how to implement it such that a coordinate pair also works. I think it's going to end up as one or the other, unless I accept *args, **kwargs and determine what the user passed in with custom logic.

def encode(self, location, grid_cells=None):
def encode(self, x, y, grid_cells=None):
def encode(self, *args, **kwargs):

[breznak]

I don't know how to implement it such that a coordinate pair also works. I think it's going to end up as one or the other,

1. is it possible to make grid-cells for nD?
2. or, encode(x,y) + override encode(position) {assert(len(position)==2; return encode(position.x, position.y)}

?

[ctrl-z-9000-times]

is it possible to make grid-cells for nD?

Not with this encoder. The code will not support it.

I think that the science is still out on whether or not real grid cells can work with arbitrary dimensions. According to the (Kropff & Treves, 2008) hypothesis, grid cells do work in any number of dimensions. They've demonstrated their model "just works" in 3D, with few or no modifications. However, these are not real grid cells.

[ctrl-z-9000-times]

or, encode(x,y) + override

This is python, there are no overloaded functions :)
Only the last / most recent definition of a function or method will count, the prior ones will be discarded.

[Thanh-Binh]

@ctrl-z-9000-times: after reading a paper at NICE workshop 2019 about using binding method for making a stable SDR, I come back to your gridcells and found that you do something like it by using Numpy.argpartition in each GC-segment.
Is my feeling right now?
Thanks

[ctrl-z-9000-times]

Thanh Binh,
I read the slides which you posted at https://discourse.numenta.org/t/biological-binding-of-sdr-in-htm/5934
And i cant figure out what they are doing. Slides #8, #9, & #10 were especially confusing. They also did not describe how the FREAK fits into their system. I assume that were was a long explanation of all of this during their presentation, which i missed.

[Thanh-Binh]

David: why do you use np.argpartition for GC encoder? Advantages?