Feeding HTM in batch mode (as opposed to line by line)

Engineering NuPIC
1

Is there a possibility to feed the HTM machinery the whole table in advance as opposed to feeding in line by line in a ‘for row in csvReader’ loop?
In my experiment, it takes about 10 ms per iteration on 32GB Linux machine which makes a 3000 rows file run for good half a minute.
Will running in a batch mode make any performance difference?

1 Like
2

No, there is no feature like you describe. HTM systems process temporal data streams step by step, not in batch. Also, 10ms per iteration seems like a good speed for an HTM system. How big is the structure? How many columns and cells per column? How big is the input space?

3

Thanks Matt.
I was trying out sets of 5,000-10,000 tuples (timestamp, lat, long).
I made sure the timestamps are evenly spaced and the lat, long -> are actually modified a little to overcome the fact that 1 lat degree is bigger that 1 long degree.
Grok didn’t work well at all so I tried with my own encoder.

4

You do not need to do this in the future as long as datetime is encoded in the input.