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

[dilpath]

A neural ODE has a NN that takes the state as input, and outputs a vector of the same size.

Do we support assigning the ODE state as the NN input, or do we need to assign each state variable individually to each NN input node? Being able to use the full state vector could be nice, but there is no ID for this yet, to use in the hybridization table.

[sebapersson]

Do we support assigning the ODE state as the NN input, or do we need to assign each state variable individually to each NN input node? Being able to use the full state vector could be nice, but there is no ID for this yet, to use in the hybridization table.

We need to assign each input separately. Still, coding the hybridization table where each state is assigned is quite straightforward as it becomes a loop over the inputs. I am also unsure if I want to start adding new util ids (as then it feels one can always motivate the need for another id set etc...)