PEtab SciML TODO

[sebapersson]

This issue acts as a TODO on things to complete for the standard:

• Add different initializationPriors. In addition the the supported priors in PEtab we should add glorot_normal, glorot_uniform as well as kaiming_normal and kaiming_uniform, where for the latter users will have to specify the gain. Need to add test cases for this.
• Add ability to set initialization for a layer via the parameters table. The following should be allowed in the parameters table: netId.layerId. Need to add test cases for this.
• Add test case where neural network parameters are constant (not estimated).
• Following PEtab YAML formatting #4, specify neural network output in the condition table. This will wait for PEtab v2 spec completion.
• Update the specification for PEtab SciML, and host it online in this repository. Usually I use Julia Documenter.jl for hosting docs, but I guess we should use something Python based for consistency?
• Update to use the code in src/python/petab_sciml for setting up the test cases.
• Add repository tests, specifically add tests to test the consistency between Lux.jl and PyTorch (this is already done in the net test-cases, but this should be refactored to a proper test directory).
• Add and test for the yaml model specification and parameter import with PyTorch?

Feel free to add any more points you might find relevant. Once the above points are addressed, I think the extension should be close to complete.

[FFroehlich]

sorry to be that guy, but in the testcases

• model_files in the yaml files needs to be a key value pairs with the model id as key
• petabEntityId and modelEntityId in the mapping tables are swapped

[sebapersson]

sorry to be that guy, but in the testcases

Thanks for noticing, at least this is easy to fix. I will look at this tomorrow. As for the model id, in this case should it match the id in the SBML file? (SBML model name)

[FFroehlich]

sorry to be that guy, but in the testcases

Thanks for noticing, at least this is easy to fix. I will look at this tomorrow. As for the model id, in this case should it match the id in the SBML file? (SBML model name)

I don't think the model id is used anywhere at this point (but could be referred to in modelEntityId), so the value is not important

[sebapersson]

@FFroehlich the formatting issues are now fixed with #12