PySD Stan Backend

[JamesPHoughton]

Lets see if this discussion can work kindof like a slack channel for the effort to add the Stan Backend. =)

[enekomartinmartinez]

Subscripts keypoints in the Python builder
Hi @hyunjimoon @Dashadower this may help you

• Subscript manager is created (has the information of all the defined subscript ranges and subranges and includes methods to manage operations between subscript variables (go to the code in https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_model_builder.py#L84 and https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/subscripts.py#L10)
• When a ElementBuilder is created the subscripts of all its components are merged to get the final subscripts of the elemen (https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_model_builder.py#L490), this allows merging the several expressions of components in one function in Python. This is just a list of the ranges that define the dimensions.
• Coordinate dictionary of the element created this means having the subscripts as {range_name: [subscripts]} (https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_model_builder.py#L493) the shape of the dictionary will be the shape of the variable
• Component subscripts dictionary is built (https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_model_builder.py#L784), same as for element but for only one component (one equation in Vensim)
• Manage mapping in a reference to a variable inside another variable equation (https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_expressions_builder.py#L1794), this means that a subscript range is mapped to another and substituted, I would recommend you to wait to implement this.
• Visit subscripts of a reference and perform subindexing if necessary (https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_expressions_builder.py#L1901 and https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_expressions_builder.py#L1910 and https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_expressions_builder.py#L2072)
• Every two expressions with different subscripts are used in an operation or function need to make the shape of both similar if they are defined with different ranges (this is done by the reshape method https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_expressions_builder.py#L65 or the reorder method https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_expressions_builder.py#L201)
• The function build element writes the equations line by line, assigning the value of each component in one subset of a predefined value if the element is defined in multiple components, is the element is defined in one, it just repeats the values along the necessary dimensions if the final subscripts don't match (https://github.com/JamesPHoughton/pysd/blob/8aa50ac3d8d937ded43d1d47ed8d459ea795228f/pysd/builders/python/python_model_builder.py#L499)

Note you need something to keep the information of the subscripts in the model and manage their information (SubscriptManager in the Python builder) and you need to track the subscripts of the expressions you are building. Then there are also some special cases where they are subscripts operations (!) at the end of the range name for example VMAX. Or components with EXCEPT keyword.

[hyunjimoon]

Related to @Dashadower and my discussion on constructing arviz inferencedata with dims, individual-population information is embedded in vensim model, so setting them (time + subscript) directly to dims will ease testing inference for a hierarchical model. For better feel for vensim subscript, I classified them into two types, space and grouped_time:
a. non_space_sharing: e.g. 92 countries for covid hospitalization. Assume no interaction between non_space_sharing countries.
b. space_sharing: e.g. disaggregation for children and adult for vaccine allocation. Interaction (transmission for covid) happens between space_sharing groups.

Thanks to Eneko's pointers,

pysd/pysd/builders/python/subscripts.py

Line 10 in 8aa50ac

class SubscriptManager:

is subscript class. During stanify refactoring could you consider importing this module as we did for AST in https://github.com/Data4DM/stanify/blob/main/stanify/builders/ast_walker.py#L6 ? @Dashadower

@enekomartinmartinez may I ask whether you have tried translating vensim subscript as xarray dims? If so, could you point us to that code?

[enekomartinmartinez]

The class you point above is the one that manages subscripts writing in the objects and merging of subscripts subranges. The xarrays, in general, are created in pysd.builders.python.python_expressions_builder. Some other xarrays are created by the External classes in the model loading o when merging data from several equations in the builder.

Depending on the expression builder, the subscripts are managed in different ways. I suggest checking first ArrayBuilder and ReferenceBuilder, as they are the most basic builders for a constants array and a reference to another variable in an equation. The ReferenceBuilder class has several methods to rename and subtracts dimensions when doing a reference, including also simple mapping management.

I hope that this helps!