Source code for message_ix_models.tests.project.test_shape

import genno
import pytest

from message_ix_models.project.shape.data import SHAPE  # noqa: F401
from message_ix_models.tools.exo_data import prepare_computer


[docs] class TestSHAPE: @pytest.mark.parametrize( "source_kw, dimensionality", ( (dict(measure="gdp", scenario="society"), {}), (dict(measure="gini", scenario="society"), {}), (dict(measure="population", scenario="all_SHAPE_SDPs"), {}), (dict(measure="urbanisation", scenario="urb_tech"), {}), pytest.param( dict(measure="foo", scenario="urb_tech"), {}, marks=pytest.mark.xfail(raises=ValueError), ), ), ) @pytest.mark.parametrize( "regions, aggregate, N_n", ( ("R12", True, 12), ("R12", False, 183), ), ) def test_prepare_computer( self, test_context, source_kw, dimensionality, regions, aggregate, N_n ): test_context.model.regions = regions source = "SHAPE" source_kw.update(aggregate=aggregate) c = genno.Computer() keys = prepare_computer(test_context, c, source, source_kw) # Key has an informative name assert source_kw["measure"] == keys[0].name # Preparation of data runs successfully result = c.get(keys[0]) # Data have the expected dimensions and size assert {"n", "y"} == set(result.dims) assert N_n <= len(result.coords["n"]) assert N_n * 14 <= result.size
[docs] @pytest.mark.parametrize( "args, size", [ (("gdp", None), 23904), (("gdp", "1.2"), 23904), (("gdp", "1.1"), 14193), (("gdp", "1.0"), 14193), (("gini", None), 9333), (("gini", "1.1"), 9333), pytest.param( ("gini", "1.0"), 9333, marks=pytest.mark.xfail( raises=genno.ComputationError, reason="File format differs, lacks 'tgt.achieved' column", ), ), (("population", None), 7968), (("population", "1.2"), 7968), (("population", "1.1"), 4731), (("population", "1.0"), 4731), (("urbanisation", None), 11001), ], ) def test_get_shape_data(test_context, args, size): test_context.model.regions = "R12" source = "SHAPE" source_kw = dict(measure=args[0], aggregate=False, interpolate=False) if args[1]: source_kw.update(version=args[1]) c = genno.Computer() keys = prepare_computer(test_context, c, source, source_kw) # Preparation of data runs successfully result = c.get(keys[0]) # Data have the expected size assert size == result.size assert {"n", "y", "SCENARIO"} == set(result.dims)