Problems in running Pavlovian Conditioning code.

[JeongInHoi]

I’m In-hoi Jeong and I’m MS-Ph.D. student of department of physics in Korea University.
I'm trying to do some simulation about pavlovian conditioning using genn simulator.
I slightly modified your Pavlovian Conditioning Model code and try to run it.
I want to set connectivity of the network manually so I changed

“e_e_pop = model.add_synapse_population("EE", "SPARSE_INDIVIDUALG", e_delay,
"E", "E",
izhikevich_stdp_model, stdp_params, stdp_init, {}, {},
"DeltaCurr", {}, {},
genn_model.init_connectivity("FixedProbability", fixed_prob_params))”
into
“e_e_pop = model.add_synapse_population("EE", "SPARSE_INDIVIDUALG", e_delay,
"E", "E",
izhikevich_stdp_model, stdp_params, stdp_init, {}, {},
"DeltaCurr", {}, {},
None)
e_e_pop.set_sparse_connections(edge_e_e[0],edge_e_e[1])
”
In here, edge_e_e is a numpy array of connectivity which was saved from previous simulation.
I also double checked about matrix type and variable type in model definition.

Unfortunately, there is an “RuntimeError: synapseUpdate.cc: 335: cuda error 700: an illegal memory access was encountered”.
I try to configure what happened and find out the error came from

“CHECK_CUDA_ERRORS(cudaMemcpyToSymbolAsync(d_mergedPresynapticUpdateGroup0, &group, sizeof(MergedPresynapticUpdateGroup0), idx * sizeof(MergedPresynapticUpdateGroup0))”

It seems like that the cudaMemcpyToSymbolAsync function try to copy data from memory to gpu but failed because of error code 700. I guess error code 700 means there is a problem with variable type.
It would be very helpful if you could provide me some advice about this failure.

Thank you very much for your time.

[tnowotny]

Hi, I think you might have broken it by passing None as the connectivity initialiser. The last argument of the connectivity initialiser is optional, so just do

e_e_pop = model.add_synapse_population("EE", "SPARSE_INDIVIDUALG", e_delay,
"E", "E", izhikevich_stdp_model, stdp_params, stdp_init, {}, {},"DeltaCurr", {}, {})

(no "None")

[JeongInHoi]

Thank you for the quick answer.
I tried as what you said but the result is the same.(same error code appears)
And according to the GeNN documentation, connectivity initialiser is given as None by default. so I think there is some other problem.

[tnowotny]

Ah ... sorry ... would have been too easy, @neworderofjamie will have a deeper look.

[neworderofjamie]

Thank you! What size_scale_factor are you passing to the get_params function?

[JeongInHoi]

size_scale_factor is 1 and weight_scale_factor is 1.0
I used 1600 excitatory neuron and 400 inhibitory neuron for the simulation.

[JeongInHoi]

I also upload simulation code for the case you need it.
https://github.com/JeongInHoi/data-for-question-genn.git

[neworderofjamie]

So I think your issue is that when you save your EI connectivity at https://github.com/JeongInHoi/data-for-question-genn/blob/main/izhikevich_pavlovian_PB_consolidation0616.py#L240 you add no_exc to the postsynaptic indices but, when you load it at https://github.com/JeongInHoi/data-for-question-genn/blob/main/common_PB_consolidation0616.py#L305-L308 you don't subtract this off again, meaning that the connectivity points to non-existent neurons, causing the error you found.

Also, your simulations are downloading the connectivity and synaptic weights and writing them to disk extremely frequently. This is very inefficient, especially for the connectivity as it does not change after the model is loaded so only needs downloading and writing to disk once!

[JeongInHoi]

Thanks a lot! Now it works and i also erase the code about downloading connectivity part:)