search

djanloo / quilt

A multiscale neural simulator
MIT License
0 stars 0 forks source link

Generation of weights and delays is senselessly slow #8

Closed djanloo closed 6 months ago

djanloo commented 7 months ago

For now the generation of weights and delays is done using Mersenne Twister algorithm through numpy. Matrices are not sparse, a lot of time and memory wasted on zero-valued entries.

djanloo commented 7 months ago

Added multithreaded SparseLogNormProjection ( 500 nanoseconds/synapse, neuron linking included) with good statisctical properties. Maybe do a Kolmogorov-Smirnov test later.

check_weights_delays

djanloo commented 6 months ago

SprseLogNorm takes a reasonable time now