[neuron-writer] how can one set the tolerance value?

[sanjayankur31]

I'm running some sims here and getting lots of nan values. Apparently, to workaround this, one has to reduce the tolerance value used in NEURON:

https://nrn.readthedocs.io/en/8.2.2/hoc/simctrl/cvode.html#CVode.atol

How would one do this while using NeuroML to generate NEURON code? (I don't think we have a way to do it yet?)

A workaround would be to generate the python script and then add a line to it, but this isn't ideal.

Edit: tasks:

• add Meta tag to LEMS schema (optional, since we don't always validate LEMS files against the schema)
• annotate Meta in jLEMS (https://docs.neuroml.org/Userdocs/LEMS_elements/Definingcomponenttypes.html#meta -> needs improvement)
• check that meta tag contents are exposed in writers
• add special case for handling cvode in neuron writer
• add example for testing, also in NEURON Showcase
• update LEMS_TEMPLATE in pyNeuroML to include Meta
• update LEMSSimulation to add method to include Meta
• regenerate docs.neuroml.org with update Meta annotation
• add special case for handling cvode in netpyne writer
• update lems simulation documentation to include netpyne options

[borismarin]

IIRC generated neuron code does not use the variable stepsize solver (cvode) - but it could: we would first need a way to specify that cvode is to be used. Not sure if that would go in the Lems definition, but jLems implements two integration methods, 4th order Runge Kutta (default) or Euler, which can be selected using the method attribute). So, we could have an additional (neuron specific) integration method (jlems would have to ignore it).

Once cvode is selected, we won't have evenly spaced points in time (i.e. time won't necessarily be an integer multiple of the dt attribute), and error tolerances should be specified instead, using eg atol as mentioned (yet another new attribute, also disregarded by jLems).

If those Simulation attrs are defined, it shouldn't be hard to update NeuronWriter.java to enable cvode and set the appropriate tolerance.

[sanjayankur31]

Thanks @borismarin , that makes sense. I think the ability to specify arbitrary simulator specific integration methods (or just a way to provide simulator specific code snippets that LEMS can ignore and "pass through" to the simulation engines) is perhaps a good general addition that we should look into. @pgleeson : what do you think?

[pgleeson]

The closest thing to this is the recommended_dt_ms and recommended_duration_ms properties: https://github.com/NeuroML/NeuroMLlite/blob/master/examples/Example4_PyNN.net.nml#L14

Though these go in the network file and are generally used if the nml file is found on its own, to help generate a useful lems file (e.g. simulating on osb v1, but copuld be used when importing the nml (not lems) into netpyne...).

I can't think of any example of a lems file that has metadata/properties with some simulator specific values. It certainly could be implemented in jnml, with jlems ignoring those options or refusing to run it, but it may be better to add general options for simulator specific info to neuromllite simulation files and more easily handle options for running them there, e.g. https://github.com/NeuroML/NeuroMLlite/blob/master/examples/Example4.py#L117, but that might require an update to the hoc/py files generated by NeuronWriter too...

[sanjayankur31]

Just making a note: we do have a "Meta" tag that seems like it was included for this purpose. The LEMS interpreter ignores it, but it should be accessible in the neuron writer;

https://github.com/LEMS/jLEMS/blob/a7a7b8662498ae55719ad22d79049077ca2ca666/src/test/resources/ex-flat.xml#L54

I'll take a look to see if we can use this to specify cvode etc.

[sanjayankur31]

LEMS/LEMS#29