Segfault in classification notebook about reservoirpy

[benureau]

Code is freshly cloned from bb00cb0

[benureau]

Ok, this is solved for me by upgrading scipy and numpy to the latest versions (1.11.2 and 1.26.0, resp.), even though reservoirpy wants scipy<=1.7.3. Consider updating the requirements or adding a warning in the installation section.

[benureau]

The segfault was triggered by this line btw:

reservoirpy/reservoirpy/nodes/readouts/ridge.py

Line 17 in bb00cb0

return linalg.solve(XXT + ridge, YXT.T, assume_a="sym")

Any ideas of issues I might encounter while using a newer scipy?

[neuronalX]

Hi @benureau, thanks for your interest in reservoirpy and for finding this issue!

Usually this kind of error, related to the regression when training the readout, could come from a bad conditioning of the matrix to be pseudo-inversed or if the size of the training data (nr of timesteps / number of reservoir dimensions) is too big.
Changing the ridge parameter could solve the first problem and reducing the data solve the second kind of problem.
But in your case this seems very related to the version of scipy in which the internal computations of linalg.solve seems to have changed.
No clear idea yet.

[PAUL-BERNARD]

Hello @benureau , thank you for your extensive report.

Even if I wasn't able to reproduce on my machine, which makes it difficult to debug, it seems that this is an issue with older scipy versions (<1.9.0), specifically on the ARM architecture.

As you mentionned, updating your scipy version should fix this. Feel free to reopen if your problem persists on recent scipy versions.