Help with segmentation fault - Community Fork of the nupic.core

Hello! I’m using the Community Fork of the nupic.core and I’ve started building my tests with the TM upon the example. After noticing my model would not forget any pattern it had learnt, making it improper for sequence learning, I’ve added the “predictedSegmentDecrement” parameter to the network. However, now I’m getting a “segmentation fault” error. The higher the value of the parameter, the earlier I get the error. For the value 0.01, I’m able to run 2121 iterations before the segfault. For the value 0.02, it’s only 426 iterations before the error occurs. It looks like a bug for me, but it could be a “newbie fault”. Could someone help me?


The TM is supposed to remember all of the things it has seen. That’s a feature not a bug.

As for the segfault, I recommend running CMake with the following 2 flags in order to do a debugging build. Debug builds have a lot of error checks which are not enabled in release mode.

I’m trying to do sequence learning / prediction, so I need the TM to forget about old sequences and learn new ones. The predictedSegmentDecrement argument should help doing that, since it’s the “amount by which segments are punished for incorrect predictions”, right?

Thank you for the reply, but I actually do not have the knowledge nor the time to acquire the knowledge to inspect the errors myself. I just wanted to report that to someone who knows the code enough to find out what could be happening.

I don’t know if this should be in another thread (@rhyolight?), but I have tried the nupic implementation today, even though it’s slower than htm.core, and I was like “No way! Not again”. I was surprised by this error:

In the above execution I was using 0.04 for predictedSegmentDecrement and It probably iterated about 300000 to 400000 times before crashing.
I’m running a really small network (64 columns, 4 cells in each column), but for millions of iterations. I’m using Ubuntu 18.04.2. I’m gonna try it on Windows 10 too, maybe I will be more lucky.