npbrain 0.2.6.5

NumpyBrain: A lightweight SNN simulation framework.

Note: NumpyBrain is a project under development. More features are coming soon. Contributions are welcome.

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 performance,you can additionally install Numba. With Numba, the speed of C or FORTRAN can be obtained in the simulation.

A variety of Python SNN simulators are available in the internet, such as Brian2, ANNarchy, NEST, etc. However, almost all of them are using the code generation approach. That is to say, the essence of these framework is let you use python scripts to control the writing of c/c++ codes. The advantage of these frameworks is obvious: they provide the easiest way to define the model (by using high-level descriptive language python), at the same time, get the fast run-time speed in the low level language (by running models in the backend c++ code). However, several drawbacks also exist:

• Any code generation framework has its own fixed templates to generate backend c++ codes. However, there will always be exceptions beyond the framework, such as the data or logical flows that the framework did not consider before. Therefore, the discrepancy emerges: If you want to generate highly efficient low-level language codes, you must provide a fixed code-generation template for high-level descriptions; Once, if you have a logic control beyond the template, you must want to extend this template. However, the extension of the framework is not a easy thing for the general users (even for mature users).

• Meanwhile, no framework is immune to errors. In Brian2 and ANNarchy, some models are wrongly coded and users are hard to correct them, such as the gap junction model for leaky integrate-and-fire neurons in Brian2 (see gapjunction_lif_in_brian2), Hodgkin–Huxley neuron model in ANNarchy (see HH_model_in_ANNarchy). These facts further point out that we need a framework that is friendly and easy for user-defines.

• Moreover, not all SNN simulations require the c++ acceleration. In code generation framework, too much times are spent in the compilation of generated c++ codes. However, for small network simulations, the running time is usually lower than that compilation time. Thus, the native NumPy codes (many functions are also written in c++) are much faster than the so called accelerated codes.

• Finally, just because of highly dependence on code generation, a lot of garbage (such as the compiled files and the link files) is left after code running, and users are hard to debug the defined models, making the model coding much more limited and difficult.

Therefore, NumpyBrain wants to provide a highly flexible and efficient SNN simulation framework for Python users. It endows the users with the fully data/logic flow control. The core of the framework is a micro-kernel, and it’s easy to understand (see How NumpyBrain works). Based on the kernel, the extension of the new models or the customization of the data/logic flows are very simple for users. Ample examples (such as LIF neuron, HH neuron, or AMPA synapse, GABA synapse and GapJunction) are also provided. Besides the consideration of flexibility, for accelerating the running speed of NumPy codes, Numba is used. For most of the times, models running on Numba backend is faster than c++ codes (see examples/benchmark).

More details about NumpyBrain please see our document.

Installation

Install NumpyBrain using pip:

$> pip install npbrain
$> # or
$> pip install git+https://github.com/chaoming0625/NumpyBrain

Install NumpyBrain using conda:

$> conda install -c oujago npbrain

Install from source code:

$> python setup.py install

The following packages need to be installed to use NumpyBrain:

• Python >= 3.5

• NumPy

• Numba

Getting started: 30 seconds to NumpyBrain

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

import numpy as np
import npbrain as nn

nn.profile.set_backend('numba')  # or "numpy"

Next, define two neuron groups:

lif1 = nn.LIF(500, noise=0.5, method='Ito_milstein')  # or method='euler'
lif2 = nn.LIF(1000, noise=1.1, method='Ito_milstein')

Then, create one Synapse to connect them both.

conn = nn.connect.fixed_prob(lif1.num, lif2.num, prob=0.2)
syn = nn.VoltageJumpSynapse(lif1, lif2, weights=0.2, connection=conn)

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

mon_lif1 = nn.StateMonitor(lif1, ['V', 'spike'])
mon_lif2 = nn.StateMonitor(lif2, ['V', 'spike'])

All above definitions help us to construct a network. Providing the name of the simulation object (for example, mon1=mon_lif1) can make us easy to access it by using net.mon1.

net = nn.Network(syn, lif1, lif2, mon1=mon_lif1, mon2=mon_lif2)

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=(lif1, 15.), report=True)

Finally, visualize the running results:

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

It shows

Define a Hodgkin–Huxley neuron model

import numpy as np
import npbrain as nn

def HH(geometry, method=None, noise=0., E_Na=50., g_Na=120., E_K=-77.,
       g_K=36., E_Leak=-54.387, g_Leak=0.03, C=1.0, Vr=-65., Vth=20.):

    var2index = {'V': 0, 'm': 1, 'h': 2, 'n': 3}
    num, geometry = nn.format_geometry(geometry)
    state = nn.initial_neu_state(4, num)

    @nn.update(method=method)
    def int_m(m, t, V):
        alpha = 0.1 * (V + 40) / (1 - np.exp(-(V + 40) / 10))
        beta = 4.0 * np.exp(-(V + 65) / 18)
        return alpha * (1 - m) - beta * m

    @nn.update(method=method)
    def int_h(h, t, V):
        alpha = 0.07 * np.exp(-(V + 65) / 20.)
        beta = 1 / (1 + np.exp(-(V + 35) / 10))
        return alpha * (1 - h) - beta * h

    @nn.update(method=method)
    def int_n(n, t, V):
        alpha = 0.01 * (V + 55) / (1 - np.exp(-(V + 55) / 10))
        beta = 0.125 * np.exp(-(V + 65) / 80)
        return alpha * (1 - n) - beta * n

    @nn.update(method=method, noise=noise / C)
    def int_V(V, t, Icur, Isyn):
        return (Icur + Isyn) / C

    def update_state(neu_state, t):
        V, Isyn = neu_state[0], neu_state[-1]
        m = nn.clip(int_m(neu_state[1], t, V), 0., 1.)
        h = nn.clip(int_h(neu_state[2], t, V), 0., 1.)
        n = nn.clip(int_n(neu_state[3], t, V), 0., 1.)
        INa = g_Na * m * m * m * h * (V - E_Na)
        IK = g_K * n ** 4 * (V - E_K)
        IL = g_Leak * (V - E_Leak)
        Icur = - INa - IK - IL
        V = int_V(V, t, Icur, Isyn)
        neu_state[0] = V
        neu_state[1] = m
        neu_state[2] = h
        neu_state[3] = n
        nn.judge_spike(neu_state, Vth, t)

    return nn.Neurons(**locals())

Acknowledgements

We would like to thank

• Risheng Lian

• Longping Liu

for valuable comments and discussions on the project.