Bug in bindings SP/Connections print

[breznak]

Reproduce: create a SP in python with a large number of inputs, print it. The output does not match.

probably some casting/integer overflow bug, either only in the print statement (low bug), or in the SP/Conn parameter.

Output

>>> from htm.bindings.sdr import SDR
>>> sp = SP([1000000],[1000])
>>> print(sp)
Spatial Pooler Connections:
Inputs (17000) ~> Outputs (1000) via Segments (1000)
Segments on Cell Min/Mean/Max 1 / 1 / 1
Potential Synapses on Segment Min/Mean/Max 17 / 17 / 17
Connected Synapses on Segment Min/Mean/Max 2 / 8.507 / 15
Synapses Dead (0%) Saturated (0%)

Inputs above incorrect.

>>> sp = SP([100],[1000])
>>> print(sp)
Spatial Pooler Connections:
Inputs (100) ~> Outputs (1000) via Segments (1000)
Segments on Cell Min/Mean/Max 1 / 1 / 1
Potential Synapses on Segment Min/Mean/Max 17 / 17 / 17
Connected Synapses on Segment Min/Mean/Max 2 / 8.507 / 15
Synapses Dead (0%) Saturated (0%)

[ctrl-z-9000-times]

I think I know why this is happening. The connections object does not know how many presynaptic inputs it has. It infers that number from the synapse data it gets. So if the spatial pooler forms zero connections from an input, then that input will not be counted.

[breznak]

. So if the spatial pooler forms zero connections from an input, then that input will not be counted.

that sounds reasonable. Should just mention in a comment "The number of inputs is infered from input-synapses formed, so it does not have to be exactly the same."

connections object does not know how many presynaptic inputs it has. It infers that number from the synapse data it gets

And let me follow up with a question, are these potential synapses? Ie, if it's not formed initially, it never would later. Or a new segment can be created on uncovered input-space if needed?

If the latter, it's ok, if the former, we should use this to check for some sensible settings. As I can form a SP 100k inputs -> 200 cols. And obviously

Spatial Pooler Connections:
Inputs (3400) ~> Outputs (200) via Segments (200)

which does not seem enough.

[ctrl-z-9000-times]

The connections object uses potential synapses

[ctrl-z-9000-times]

check for some sensible settings.

There is an example of checking for sane connections in the branch "columnPooler2" but I don't know how applicable that code is to the Spatial Pooler.