Tutorial 1

In this tutorial, we are going to build an SNN capable of classifying MNIST by copying the weights obtained by training the following simple ANN using TensorFlow:

Using GeNN for spike-based machine learning.svg

Clearly, this is far from a state of the art architecture, but it still achieves 97.6% accuracy on MNIST.

Install

Download wheel file

[1]:

if "google.colab" in str(get_ipython()):
    !gdown 1qlNCa_WT7sOmifYjPgGyggqUGsPiHn5y
    !pip install pygenn-5.3.0-cp312-cp312-linux_x86_64.whl
    %env CUDA_PATH=/usr/local/cuda

    !rm -rf /content/ml_genn-ml_genn_2_4_0
    !wget https://github.com/genn-team/ml_genn/archive/refs/tags/ml_genn_2_4_0.zip
    !unzip -q ml_genn_2_4_0.zip
    !pip install ./ml_genn-ml_genn_2_4_0/ml_genn

Downloading...
From: https://drive.google.com/uc?id=1QF6eMWoqmOehbzXNSUbrImyBo0dTbv6J
To: /content/pygenn-5.2.0-cp311-cp311-linux_x86_64.whl
100% 8.60M/8.60M [00:00<00:00, 138MB/s]
Processing ./pygenn-5.2.0-cp311-cp311-linux_x86_64.whl
Requirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.11/dist-packages (from pygenn==5.2.0) (2.0.2)
Requirement already satisfied: psutil in /usr/local/lib/python3.11/dist-packages (from pygenn==5.2.0) (5.9.5)
Requirement already satisfied: setuptools in /usr/local/lib/python3.11/dist-packages (from pygenn==5.2.0) (75.2.0)
pygenn is already installed with the same version as the provided wheel. Use --force-reinstall to force an installation of the wheel.
env: CUDA_PATH=/usr/local/cuda
--2025-04-25 14:56:29--  https://github.com/genn-team/ml_genn/archive/refs/tags/ml_genn_2_3_1.zip
Resolving github.com (github.com)... 20.205.243.166
Connecting to github.com (github.com)|20.205.243.166|:443... connected.
HTTP request sent, awaiting response... 302 Found
Location: https://codeload.github.com/genn-team/ml_genn/zip/refs/tags/ml_genn_2_3_1 [following]
--2025-04-25 14:56:29--  https://codeload.github.com/genn-team/ml_genn/zip/refs/tags/ml_genn_2_3_1
Resolving codeload.github.com (codeload.github.com)... 20.205.243.165
Connecting to codeload.github.com (codeload.github.com)|20.205.243.165|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: unspecified [application/zip]
Saving to: ‘ml_genn_2_3_1.zip.1’

ml_genn_2_3_1.zip.1     [ <=>                ] 689.46K  --.-KB/s    in 0.06s

2025-04-25 14:56:29 (12.0 MB/s) - ‘ml_genn_2_3_1.zip.1’ saved [706004]

Processing ./ml_genn-ml_genn_2_3_1/ml_genn
  Preparing metadata (setup.py) ... done
Requirement already satisfied: pygenn<6.0.0,>=5.1.0 in /usr/local/lib/python3.11/dist-packages (from ml_genn==2.3.1) (5.2.0)
Collecting enum-compat (from ml_genn==2.3.1)
  Downloading enum_compat-0.0.3-py3-none-any.whl.metadata (954 bytes)
Requirement already satisfied: tqdm>=4.27.0 in /usr/local/lib/python3.11/dist-packages (from ml_genn==2.3.1) (4.67.1)
Requirement already satisfied: deprecated in /usr/local/lib/python3.11/dist-packages (from ml_genn==2.3.1) (1.2.18)
Requirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.11/dist-packages (from pygenn<6.0.0,>=5.1.0->ml_genn==2.3.1) (2.0.2)
Requirement already satisfied: psutil in /usr/local/lib/python3.11/dist-packages (from pygenn<6.0.0,>=5.1.0->ml_genn==2.3.1) (5.9.5)
Requirement already satisfied: setuptools in /usr/local/lib/python3.11/dist-packages (from pygenn<6.0.0,>=5.1.0->ml_genn==2.3.1) (75.2.0)
Requirement already satisfied: wrapt<2,>=1.10 in /usr/local/lib/python3.11/dist-packages (from deprecated->ml_genn==2.3.1) (1.17.2)
Downloading enum_compat-0.0.3-py3-none-any.whl (1.3 kB)
Building wheels for collected packages: ml_genn
  Building wheel for ml_genn (setup.py) ... done
  Created wheel for ml_genn: filename=ml_genn-2.3.1-py3-none-any.whl size=132274 sha256=1e56cd8677025ed90c01f4234ca2b54a07acaa46d867118bc3ed1132eac4f9c8
  Stored in directory: /tmp/pip-ephem-wheel-cache-09fb_emv/wheels/ad/47/4c/9f1426577cf209699f869217d188ff31354668a5b50067c04b
Successfully built ml_genn
Installing collected packages: enum-compat, ml_genn
Successfully installed enum-compat-0.0.3 ml_genn-2.3.1

Download pre-trained weights

[2]:

!gdown 1cmNL8W0QZZtn3dPHiOQnVjGAYTk6Rhpc
!gdown 131lCXLEH6aTXnBZ9Nh4eJLSy5DQ6LKSF

Downloading...
From: https://drive.google.com/uc?id=1cmNL8W0QZZtn3dPHiOQnVjGAYTk6Rhpc
To: /content/weights_0_1.npy
100% 402k/402k [00:00<00:00, 141MB/s]
Downloading...
From: https://drive.google.com/uc?id=131lCXLEH6aTXnBZ9Nh4eJLSy5DQ6LKSF
To: /content/weights_1_2.npy
100% 5.25k/5.25k [00:00<00:00, 20.2MB/s]

Install MNIST package

[3]:

!pip install mnist

Collecting mnist
  Downloading mnist-0.2.2-py2.py3-none-any.whl.metadata (1.6 kB)
Requirement already satisfied: numpy in /usr/local/lib/python3.11/dist-packages (from mnist) (2.0.2)
Downloading mnist-0.2.2-py2.py3-none-any.whl (3.5 kB)
Installing collected packages: mnist
Successfully installed mnist-0.2.2

Build model

Import standard modules and required mlGeNN classes

[4]:

import mnist
import numpy as np

from ml_genn import InputLayer, Layer, SequentialNetwork
from ml_genn.compilers import InferenceCompiler
from ml_genn.neurons import IntegrateFire, IntegrateFireInput
from ml_genn.connectivity import Dense
from ml_genn.callbacks import SpikeRecorder

Because our network is entirely feedforward, we can define it as a SequentialNetwork where each layer is automatically connected to the previous layer. We are going to convert MNIST digits to spikes by simply treating the intensity of each pixel (multiplied by a scaling factor) as the input to an integrate-and-fire neuron. For our hidden and output layers we are going use very simple Integrate-and-Fire neurons which best match the transfer function of the ReLU neurons our ANN was trained with. Finally, we are going to read classifications from our output layer by simply counting spikes.

[5]:

# Create sequential model
network = SequentialNetwork()
with network:
    input = InputLayer(IntegrateFireInput(v_thresh=5.0), 784)
    Layer(Dense(weight=np.load("weights_0_1.npy")),
          IntegrateFire(v_thresh=5.0))
    output = Layer(Dense(weight=np.load("weights_1_2.npy")),
                   IntegrateFire(v_thresh=5.0, readout="spike_count"))

In mlGeNN, in order to turn an abstract network description into something that can actually be used for training or inference you use a compiler class. Here, as we are performing inference, we use the InferenceCompiler and specify batch size and how many timesteps to evaluate each example for.

[6]:

compiler = InferenceCompiler(dt=1.0, batch_size=128,
                             evaluate_timesteps=100)
compiled_net = compiler.compile(network)

We can then load the MNIST testing data and evaluate the compiled pre-trained network on it. The normal grayscale values of the MNIST dataset are passed into the input layer, scaled by a factor (here 0.01) which is translated into spikes by the IntegrateFireInput layer. The testing labels are passed as a simple array of of 0 to 9, corresponding to the default SparseCategorical loss type.

[7]:

# Load testing
mnist.datasets_url = "https://storage.googleapis.com/cvdf-datasets/mnist/"
testing_images = np.reshape(mnist.test_images(), (-1, 784))
testing_labels = mnist.test_labels()

with compiled_net:
    # Evaluate model on numpy dataset
    metrics, _ = compiled_net.evaluate({input: testing_images * 0.01},
                                       {output: testing_labels})