npbrain 0.2.0

NumpyBrain: A lightweight micro-core SNN simulation framework.

Why to use NumpyBrain

NumpyBrain is a microkernel framework for SNN (spiking neural network) simulation purely based on native python. It only relies on NumPy. However, if you want to get faster CPU performance, or run codes on GPU, you can additionally install Numba. With Numba, the speed of C or FORTRAN can be gained in the simulation.

A variety of Python SNN simulators are available in the internet, such as Brain2, ANNarchy, NEST, etc. However, several reasons motivate us to write a NumPy-based simulator.

• First of all, using these simulators, a lot of garbage files are left after code compiling and running. In Brain2, such annoying rubbish can even accumulate to several GB.

• Second, the inner run-time mechanism is difficult to understand. The essence of these framework is that let you use python scripts to control the writing of c++/CUDA codes. Thus, if the users are not familiar with c++/CUDA codes and don’t understand the inner mechanism of the framework, the data and logic flow control will be very complex and incomprehensible.

• Because of this, under these frameworks, on one hand, the codings of some models are weird, for example the Gap Junction model in Brian2 (which dramatically different from other kinds of synapses); on the other hand, some models are wrongly coded and are hard to correct, such as the Gap Junction model for LIF (leaky integrate-and-fire) neurons in Brian2 (see some code.py), Hodgkin–Huxley neuron model in ANNarchy (see some code.py).

Therefore, NumpyBrain want to provide a highly flexible SNN simulation framework for Python users. It endows the users with the fully data/logic flow control. Its design overcomes the defects of other simulators, and are guided by the following principles:

• Plug and play. No garbage file will be generated and left after any code-running. Just, use or not use.

• Modularity. A network can be broken down into various neurons and synapses. To inspect the inner dynamical structure of these elements, we need the Monitor to record the running trajectory for each object. In NumpyBrain, there are only these three kinds of objects. Such objects can be plugged together almost arbitrarily (only with few restrictions) to form a new network.

• Easy extensibility. For each kind of object, new models (neurons or synapses) are simple to add, and existing models provide ample examples.

• User friendliness. The data flow in each object is transparent, and can be easily controlled by users. Users can define or modify the data or logical flow by themselves according to need.

Getting started: 30 seconds to NumpyBrain

First of all, before import the package, you can set the numerical backend you prefer:

from npbrain import profile
profile.set_backend('numba')

Then, import the package all. Later you will find this makes you easily use all the pre-defined models:

import npbrain.all as nn

Next, all we need to do is to define neurons and synapses needed by our network. Let’s define two LIF neuron groups and one Synapse to connect them both.

lif1 = nn.LIF(500, noise=1.1)
lif2 = nn.LIF(1000, noise=1.1)
syn = nn.NormalSynapses(lif1, lif2, 0.2, {'method': 'fixed_prob', 'prob': 0.1})

In order to inspect the dynamics of two LIF neuron groups, we use StateMonitor to record the membrane potential and use SpikeMonitor to receive the spiking events.

mon_lif1_v = nn.StateMonitor(lif1)
mon_lif1_sp = nn.SpikeMonitor(lif1)
mon_lif2_v = nn.StateMonitor(lif2)
mon_lif2_sp = nn.SpikeMonitor(lif2)

All above definitions help us to construct a network. Providing the name of the simulation object (for example, mon1=mon_lif1_v) can make us easy to access it by using net.mon1.

net = nn.Network(lif1, lif2, syn, mon1=mon_lif1_v, mon2=mon_lif1_sp,
                 mon3=mon_lif2_v, mon4=mon_lif2_sp)

We can simulate the whole network just use .run(duration) function. Here, we set the inputs of lif1 object to 15., and open the report mode.

net.run(duration=100, inputs=15., receiver=lif1, report=True)

Finally, visualize the running results:

import matplotlib.pyplot as plt

fig, gs = nn.vis.get_figure(n_row=2, n_col=1, len_row=3, len_col=8)
nn.vis.plot_potential(net.run_time, net.mon1, ax=fig.add_subplot(gs[0, 0]))
nn.vis.plot_raster(net.mon2, ax=fig.add_subplot(gs[1, 0]))
plt.show()

It shows

Documentation

Available online documents: latest docs.

Available offline PDF: latest PDF.

Installation

Install NumpyBrain using pip:

$> pip install npbrain
$> # or
$> pip install git+https://github.com/oujago/NumpyBrain

Install NumpyBrain using conda:

$> conda install -c oujago npbrain

Install from source code:

$> python setup.py install

Dependency

The following packages need to be installed to use NumpyBrain:

• Python3

• NumPy

Recommended:

• Numba