pypet PicklingError about neurolib HOT 4 CLOSED

Dear Valerii,

Thanks for your interest in neurolib!
Unfortunately, due to some breaking changes related to forking a process, python3.8 is not currently supported on macOS systems. Our dependency on the pypet package and its multiprocessing scheme is the culprit. We thought of migrating for some different solutions, but for now, please use a lower version of Python. Pythons 3.6 and 3.7 are tested regularly and working as they should on MacOS. Myself, I am running 11.3 and python 3.7.10 for development.

Just as a good measure, I successfully ran the code you posted with python 3.7.10 with a fresh install of neurolib in a temporary virtual environment:

I hope this intermediate solution works for you and you are able to accommodate python 3.7 for now :)

Thanks again and happy brain simulating!

from neurolib.

Hey,

no problem, glad to help:)
Well, yes. For this particular use case, the single processing will always be faster, since one simulation of the ALN 1 node model for 2seconds can be done in a tiny fraction of a second, and then the overhead caused by forking a process and then waiting for them to synchronize is much much larger than the simulation itself. The more complex the model is (therefore, the more a single simulation would take), the more multiprocessing makes sense.

You can try to do the same thing with the brain network instead of one node, e.g. do something like:

# make 20-node brain network
model = ALNModel(Cmat=np.random.rand(20,20))
# simulate for one minute
model.params["duration"] = 60.0 * 1e3 # in ms

to create a brain network with 20 nodes, connected randomly. Now the simulation of a single time will be significantly longer, and you'll benefit from multiprocessing on exploration. In general, multiprocessing always comes with an overhead of forking a process and waiting for them to sync, so it is not always the best option to use it.

By the way, I am also using an M1 machine and instead of the non-existing gcc for arm64, I didn't encounter any problems so far:)

Best,
N.

from neurolib.

Dear @jajcayn,

Thank you for your quick reply and your help!

I've changed python to 3.7.10, and everything works fine, although I've noticed that with multiproc=False it works much faster. Maybe this is because I am using Apple M1...

from neurolib.

Hey,

no problem, glad to help:)
Well, yes. For this particular use case, the single processing will always be faster, since one simulation of the ALN 1 node model for 2seconds can be done in a tiny fraction of a second, and then the overhead caused by forking a process and then waiting for them to synchronize is much much larger than the simulation itself. The more complex the model is (therefore, the more a single simulation would take), the more multiprocessing makes sense.

You can try to do the same thing with the brain network instead of one node, e.g. do something like:

# make 20-node brain network
model = ALNModel(Cmat=np.random.rand(20,20))
# simulate for one minute
model.params["duration"] = 60.0 * 1e3 # in ms

to create a brain network with 20 nodes, connected randomly. Now the simulation of a single time will be significantly longer, and you'll benefit from multiprocessing on exploration. In general, multiprocessing always comes with an overhead of forking a process and waiting for them to sync, so it is not always the best option to use it.

By the way, I am also using an M1 machine and instead of the non-existing gcc for arm64, I didn't encounter any problems so far:)

Best,
N.

Alright! Thank you very much for the explanation!

I think I will close the issue because everything works fine on my computer.

Thank you for your help!

from neurolib.