policyengine / policyengine-core Goto Github PK

When trying to instantiate a Dataset object (in the same way I was successfully doing before the model-name-change process was undertaken), I get the following error message:

Traceback (most recent call last):
.
.
.
  File "//Users/mrh/anaconda3/envs/policyengine-us/lib/python3.9/site-packages/policyengine_core/data/dataset.py", line 41, in __init__
    if not self.folder_path.exists():
AttributeError: 'str' object has no attribute 'exists'

The error message says it all: a string (folder_path) does not have an exists method.

If you are trying to check the existence of the folder_path, then try something like this:

if not os.path.exists(self.folder_path):

We are on

openfisca-core is now on numpy >=1.24.2, <1.25. They updated to >=1.24.2 in openfisca/openfisca-core#1181 (which has many changes due to code reformatting), and limited to <1.25 in openfisca/openfisca-core#1184.

Type hints, as well as being good practice, enable IDEs like Visual Studio Code to provide relevant information about properties at the time of writing. We should annotate all non-inferable types in the codebase to aid this.

For example, PolicyEngine/policyengine-canada#313 (review) refers to a tax credit with a maximum amount that partially phases out over a dollar range, as does the Guaranteed Income for the 21st Century.

We can use add instead

cc @MaxGhenis

Following warning is generated when running tests on GitHub:

policyengine_us/tests/microsimulation/test_microsim.py: 3 warnings
  /opt/hostedtoolcache/Python/3.9.17/x64/lib/python3.9/site-packages/policyengine_core/enums/enum.py:74: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison
    if isinstance(array == 0, bool):

Code:

We should allow country packages to define their own time periods. For example, the UK fiscal year begins on the 1st of April. I think the cleanest implementation would involve the Period class having period-specific features removed, and adding base classes Day, Month, Year and Eternity inherit from Period.

PRs #127 and/or #128 fixed bugs that affected UK microsimulation accuracy with respect to uprating and reforms, but caused the US one to break entirely (PolicyEngine/policyengine-us#3311).

This is now the latest stable version of Python

I'd like to enforce via CI all pdf references in parameters including a page number via the .pdf#page=x URL.

Doesn't occur when reinitialising from a new MicroDataFrame. Found in this notebook.

If a scale parameter has a threshold_variable metadata field (#83), we could automatically calculate scale parameters.

For example, if the tax.mtr_schedule parameter had a threshold_variable: taxable_income metadata field, we could condense this formula:

taxable_income = tax_unit("taxable_income", period)
tax = parameters(period).tax.mtr_schedule.calc(taxable_income)

into this:

tax = parameters(period).tax.mtr_schedule.calc()

This further condenses the proposal in #84.

Currently, yaml tests without an input throw a cryptic error message:

AttributeError: ‘NoneType’ object has no attribute ‘items’

We could throw a more informative error message, but I think assuming input: {} when no input is provided will be easier.

Currently the model returns an error

Some users may want to make a venv (and we may want to suggest this), so ignoring contents will avoid git issues there.

For posterity, steps would be something like

pyenv shell 3.9.16 
python -m venv venv  

cc @ibushong

For example PolicyEngine/policyengine-us#3409 introduced invalid parameter YAML (see PolicyEngine/policyengine-us#3417), but didn't break CI

The API did catch it though: https://github.com/PolicyEngine/policyengine-api/actions/runs/7162451862/job/19499442307?pr=964

For example see PolicyEngine/policyengine-us#2815 (comment), where

adds = "gov.states.in.tax.income.credits.refundable"

breaks due to in being a keyword.

In advance of addressing paths with these keywords in general, we could parse them in the adds calls to replace all .x with ['x'] when parsing it.

Currently, Enum variables often lose their encodings when mapping between entities.

As a step toward #37, per #37 (comment)

To condense this:

taxable_income = tax_unit("taxable_income", period)
tax = parameters(period).tax.mtr_schedule.calc(taxable_income)

to this:

tax = parameters(period).tax.mtr_schedule.calc("taxable_income")

The Simulation class in Core is useful as a base class, but there is a special case of Simulation which deserves its own interface: simulations on weighted survey microdata. We have an implementation in OpenFisca-Tools which we should port over to Core.

Upgrading wheel, setuptools, and pip before installing other requirements can avert some environment issues. We can do this with a separate requirements.txt file that does something like pip install -U wheel setuptools pip (but with pins) and adding something like make setup as a pre-install step.

Thanks @ibushong for the tip here!

For example, suppose we're using adds to create a variable representing a state's refundable credits based on a policy parameter, and the state only has one refundable credit in some years, zero in others. It currently breaks when trying to sum from the empty list. @PavelMakarchuk

Magic methods are what enable machine learning libraries like TensorFlow or PyTorch to compile Python expressions of Tensors into optimised instructions. Can we do the same with PolicyEngine country packages?

Python 3.6 is at end of life

Should be the answer to "Is Y X", and in the documentation as best practice.

For example, PolicyEngine/policyengine-us#2306 involves a nh_total_exemptions variable which sums several individual exemption variables using adds=[], and one of those is a nh_base_exemption which is a flat amount. So it'd be easier to avoid a formula just to return a parameter value.

To enable this:

description: NYC provides a household credit of this fixed amount to single filers, based on recomputed federal AGI.
metadata:
  label: NYC Household Credit Fixed Amount for single filers
  type: single_amount
  threshold_unit: currency-USD
  amount_unit: currency-USD
  threshold_variable: adjusted_gross_income # PROPOSAL
  reference:
    - title: Instructions for Form IT-201
      href: https://www.tax.ny.gov/pdf/2022/printable-pdfs/inc/it201i-2022.pdf#page=17
values:
  - threshold:
      2022-01-01: 10_000
    amount:
      2022-01-01: 15
  - threshold:
      2022-01-01: 12_500
    amount:
      2022-01-01: 10
  - threshold:
      2022-01-01: 12_501
    amount:
      2022-01-01: 0

cc @leogoldman

Like sum_of_variables does.

Following these instructions: https://policyengine.github.io/policyengine-core/usage/cli.html#data

(policyengine) maxghenis@MacBook-Air-3 policyengine-us % policyengine-core data datasets list -c policyengine_us
Traceback (most recent call last):
  File "/Users/maxghenis/miniconda3/envs/policyengine/bin/policyengine-core", line 8, in <module>
    sys.exit(main())
  File "/Users/maxghenis/miniconda3/envs/policyengine/lib/python3.9/site-packages/policyengine_core/scripts/policyengine_command.py", line 128, in main
    return sys.exit(main(parser))
  File "/Users/maxghenis/miniconda3/envs/policyengine/lib/python3.9/site-packages/policyengine_core/scripts/run_data.py", line 36, in main
    print(dataset_summary(datasets))
  File "/Users/maxghenis/miniconda3/envs/policyengine/lib/python3.9/site-packages/policyengine_core/scripts/run_data.py", line 12, in dataset_summary
    years = list(sorted(list(set(sum([ds.years for ds in datasets], [])))))
  File "/Users/maxghenis/miniconda3/envs/policyengine/lib/python3.9/site-packages/policyengine_core/scripts/run_data.py", line 12, in <listcomp>
    years = list(sorted(list(set(sum([ds.years for ds in datasets], [])))))
AttributeError: type object 'CPS_2020' has no attribute 'years'

To avoid issues like in this PR, which had a missing amount that went undiagnosed.

We should port the defined-for logic from OpenFisca-Tools, making improvements where we can.

Per PolicyEngine/policyengine-uk#645, pytest version requirements are causing conflicts.

We restrict the version range in a way that mirrors openfisca-core. Can we uncap the version?

Not high priority and IIUC it's blocked by a pytest issue (#32), but dropping here for the future

Pending #68, which would bring the same functionality via adds

TWO THINGS: what seems to be a bug and a feature request:

BUG:
In the Dataset class, exists is a property not a method.
So, should the error-causing line be:

with h5py.File(file, "a" if file.exists else "w") as f:

ERROR MESSAGE:

vdset.save_dataset(vdset)
  File "//Users/mrh/anaconda3/envs/policyengine-us/lib/python3.9/site-packages/policyengine_core/data/dataset.py", line 172, in save_dataset
    with h5py.File(file, "a" if file.exists() else "w") as f:
AttributeError: 'str' object has no attribute 'exists'

FEATURE REQUEST:
Also, could you add an optional argument to the save_dataset method that allows use of a file path that differs from self.file_path?

Based on the cached baseline

cc @MichaelSnowden @ibushong

A common task analysts need to carry out is to understand why reforms impact subsections of society. It'd be really helpful for a (e.g.) streamlit-based chatbot to be able to answer questions about a reform, for example:

Pull out a household that benefits from the reform in the code box above.

OK, I picked household #2230.

Explain why they benefited?

Their net income increased by $X. It looks like this was because of their benefits, which in turn was because of their SNAP.

Looking through the package contents, there's a lot of legacy code that is unused in the latest version and exists as a result of previous evolutions of the codebase. For example, the Scenario interface that was succeeded by SimulationBuilder. I think removing these features will make the package more maintainable and help to reduce bugs.

The property is empty but should contain all variables which were specified in the test.

We're seeing a common pattern in modelling legislation: rules that take the form:

variable X = what you would get if parameter Y were instead Z

Inside variable formulas, we should be able to 'branch off' a new simulation, modifying parameters and calculating variables. We already have an implementation in OpenFisca Tools we can port over, though we should investigate to see if there's a cleaner way now that we can edit Core directly in this fork.

Often references apply to children, but not labels. We should allow this propagation to distinguish between them.

Currently, references are a union of different formats: strings, tuples and lists. We should define a concrete Reference class, with a defined JSON translation, to make detailed references machine-friendly.

We've had several undetected crashes arising from formulas that mix up entity types, resulting in this type of error:

ValueError: operands could not be broadcast together with shapes (100000,) (279932,)

For example, PolicyEngine/policyengine-us#3298

Let's test for this centrally.

@martinholmer could you share the command you executed in that issue to throw the error?

numpy is now on version 1.23, but we cap it, following openfisca-core:

Definition periods in PolicyEngine Core are fixed: each variable must have exactly one time period for which it can have distinct values. Policy in country models often involves mid-year parameter changes or other ways that more flexible time period management might be useful.

Core has some shortcuts for this (e.g. set_input_divide, options=[ADD, DIVIDE]) but a more elegant solution to reduce the development burden on country models would just be to simplify all this into more autonomous Simulation behaviour: e.g. if I said my employment income was £30,000 per year, calculate my income tax in March, we really shouldn't need any country-specific logic describing how the time periods interact with each other here. The only thing we need to specify at the country level is whether each variable is a stock or a flow, because this implies different behaviour is correct:

• wealth in 2022 is 450k => wealth in 2022-01 is 450k
• income in 2022 is 30k => income in 2022-01 is 2.5k