Unable to run cpp_standalone in multiple test files

[zifeo]

I have multiple test files, each containing one single network that is explicitly build:

set_device('cpp_standalone')
# ...
net = Network()
net.add(...)
net.run(1 * s)

When the tests are run individually (pytest), everything is fine. However when running the whole folder I am getting the following, even though they are separated networks:

RuntimeError: The network has already been built and run before. Use set_device with build_on_run=False and an explicit device.build call to use multiple run statements with this device.

Is there a way of resetting the global shared state? Moreover is there a way of getting rid of the output folder that is created (codegen) when we get a crash (or a test failure), maybe using a flag?

[mstimberg]

You'll have to use:

device.reinit()
device.activate()

between the runs (see http://brian2.readthedocs.io/en/stable/user/computation.html#multiple-builds).

Moreover is there a way of getting rid of the output folder that is created (codegen) when we get a crash (or a test failure), maybe using a flag?

No, there's no mechanism for that, you'll have to delete the folder manually.

[zifeo]

Thanks!

[zifeo]

I'm getting randomly some ValueError: All objects need to have unique names in standalone mode, the following name(s) were used more than once while running tests based on cpp_standalone (the network is constructed manually). Is there a delay for the reinit or activate?

[mstimberg]

Hmm, this is odd. The name system is independent of the device. We keep track of objects via weak references and this system can be somewhat unpredictable because those references are freed by the garbage collection process. But this should only lead to unnecessarily creating non-redundant names (e.g. naming a NeuronGroup neurongroup_2 even though no other NeuronGroup exist) -- I don't see how this could lead to duplicate names. When you say "the network is constructed manually" you mean that you use net = Network(...); net.run(...)? What are these names that are used more than once? It should not be possible to add objects more than once to a Network, but maybe there are corner cases where we don't catch this problem... Do you maybe add subgroups to the network?

[zifeo]

I've tried to trigger garbage collection in between without success. Yes, I meant using explicit calls to Network(...) and then add objects. I'm using the same names for the recreated objects I add into the recreated network, no subgroup. Maybe self.networks is not freed "fast enough" in some cases? It seems to work consistently under Python 3.6 with time.sleep(0.25) and require more sleeping time under Python 3.5. The check concerned is https://github.com/brian-team/brian2/blob/master/brian2/devices/cpp_standalone/device.py#L1129-L1130.

[mstimberg]

I see. The way we handle these things is not great, I think we have to do something about it... We fill self.networks when you do device.build, by looking at all Network instances that exist (as tracked by weak references). I think in your case there's still the old reference around at that point. Where did you place the garbage collection? I think it should do the job if you call it after the device.reinit().

But this is just a workaround, this is a bug that needs to be fixed properly in some way.

[zifeo]

I tried the following, maybe I had bad luck with it.

device.reinit()
gc.collect()
device.activate()