Add synapse population from DENSE_INDIVIDUALG to SPARSE

[riccardopignari]

Hi everyone!
I'm trying to make a small Python script using pygenn about synaptic connections between two populations of neurons. So far I've successfully implemented a DENSE_INDIVIDUALG connection using numpy (code below). I would like to know if it's possible to get an equivalent connection with a SPARSE mode.

The installed PyGeNN version is pygenn-4.7.1-cp36-cp36m-linux_x86_64.whl.

What could be the equivalent of the code below with sparse connection?

from pygenn import GeNNModel
import numpy as np
import os
import matplotlib.pyplot as plt

################################
# ##### Model definition ##### #
################################
# ##### Model setting ##### #
model = GeNNModel(precision='float', model_name=f'mainTest', backend='SingleThreadedCPU')
model.dT = 1.0


# ##### Neuron Parameters ##### #
paramLif = {
    'C': 0.25,  # nF
    'TauM': 20.0,  # ms
    'Ioffset': 0.3,  # nA
    'Vrest': -65.0,  # mV
    'Vthresh': -50.0,  # mV
    'Vreset': -70.0,  # mV
    'TauRefrac': 2.0,  # ms
}
varLif = {
    'V': paramLif['Vrest'],  # mV
    'RefracTime': 0.0,  # ms
}


# ##### Neuron population ##### #
popStim = model.add_neuron_population(
    pop_name=f'stim',
    num_neurons=3,
    neuron='PoissonNew',
    param_space={'rate': 20}, var_space={'timeStepToSpike': 0}
)

popVar = model.add_neuron_population(
    pop_name=f'var',
    num_neurons=2,
    neuron='LIF',
    param_space=paramLif, var_space=varLif
)
popVar.spike_recording_enabled = True


# ##### Synaptic connection ##### #
##### Stimulus excitation #####
synapsesExcit = np.array([
    [1.2, 1.3],
    [0.3, 1.7],
    [1.4, 0.2]
])
model.add_synapse_population(
    pop_name=f'stimExcit', matrix_type='DENSE_INDIVIDUALG',
    source=popStim, target=popVar,
    postsyn_model='ExpCurr',
    ps_param_space={"tau": 5.0}, ps_var_space={},
    delay_steps=2,
    w_update_model='StaticPulse',
    wu_param_space={}, wu_var_space={'g': synapsesExcit.flatten()},
    wu_pre_var_space={}, wu_post_var_space={}
)
##### Internal inhibition #####
synapsesInhib = np.array([
    [-1.8, -0.3],
    [-0.1, -1.5],
    [-1.7, -0.2]
])
model.add_synapse_population(
    pop_name=f'internalInhib', matrix_type='DENSE_INDIVIDUALG',
    source=popStim, target=popVar,
    postsyn_model='ExpCurr',
    ps_param_space={"tau": 5.0}, ps_var_space={},
    delay_steps=2,
    w_update_model='StaticPulse',
    wu_param_space={}, wu_var_space={'g': synapsesInhib.flatten()},
    wu_pre_var_space={}, wu_post_var_space={}
)


##########################
# ##### Simulation ##### #
##########################
timeSimulation = 1  # s
timeSteps = int(timeSimulation * 1e3)
model.build()
model.load(num_recording_timesteps=timeSteps)

while model.timestep < timeSteps:
    model.step_time()
model.pull_recording_buffers_from_device()

times, index = popVar.spike_recording_data
records = [[] for _ in range(2)]
for t, i in list(zip(times, index)):
    records[i].append(t)

plt.figure()
plt.eventplot(records)
plt.show()

os.system('rm -r *_CODE*')

[tnowotny]

You can set the matrix_type to SPARSE_INDIVIDUALG. Then, have a look here: https://genn-team.github.io/genn/documentation/4/html/d5/d39/subsect34.html for some documentation.
Probably you will want to use pygenn.SynapseGroup.set_sparse_connections to set your connections and synaptic weights.

[tnowotny]

PS: Is there a specific reason why you are not using GeNN 5?

[riccardopignari]

Thanks for the reply. I have seen the documentation, but I can't seem to initialize the synaptic population properly. Since I have layers with a very large number of neurons and sparse connections, creating a numpy 86400^2 matrix is ​​not feasible. What I was trying to do is a connection like this:

synapsesExcit = [
    [idxNeurLayer1, idxNeurLayer2, weight1],
    [idxNeurLayer1, idxNeurLayer2, weight2],
    [idxNeurLayer1, idxNeurLayer2, weight3],
    ...
]

The reason behind why I still use an old version of GENN is because I develop on a remote server environment where I have no ability to modify.