Neural Network Simulator for Blender2.5 (PyPy optimized)

Compares performance of Python vs PyPy

Rendering of simulation

Source code

Although its not tightly integrated with Blender yet (Blender is just used for rendering), i do plan to integrate many features to tweak the network (running in realtime) and interact with it from Blender once the python API is ready (once the script-links replacement comes in).

Probably be extra-special fast if you tuned it with cython.

Pretty cool visuals on the full net rendering.

very interesting. What purpose do you train the network for?

pildanovak, i’m planning to train the network for realtime musical input/output. A microphone will be sampled and trigger spikes in the network, then the FFT of the network will generate or pick notes to play. The network is generic so it could be used for other things as well, the strength of the recurrent model is that it can learn temporal sequences and timing.

Updated the Blender exporter with flashing dendrites and thick to thin dendrites:

Have you tried other optimization method for Python such as psyco? Is PyPy the fastest option? I am interested to see how they compare.

regarding animation, it would be nice to see impulses sent through the network, rather than seeing a neuron flash with all the connections at once?

mpan3, i have used Psyco on this simulator before, and found its not much help, Psyco accelerates integers better than floating point. The big speed up is to JIT the neuron iteration loop, i have the code in place for that, i am just waiting for 64bit JIT to be finished in PyPy before testing it.

pildanovak, i’m planing to have the impulses, and animated growth of dendrites as well. Just got stuck on the python API recently, anybody know how to add a Hook modifier and assign it to a vertex group on a curve from python?