def test_sum_join_multiple_datasets_undermatching():
workers = deepcopy(WORKERS_DATASET_DEF)
workers.name = 'WORKERS'
more = deepcopy(MOREWORKERS_DATASET_DEF)
more._add_var(
v('joined', 'sum(join(name == WORKERS.name, WORKERS.max_production))'))
sim = run_sim(
[],
datasets=[more, workers],
)
# Second row of MOREWORKERS doesn't match any rows in WORKERS, hence sum is 0
sim.assert_values_match(
'MOREWORKERS.joined',
[200, 0, 100],
t=1,
)
def test_time_indexing_dataset_vars():
db = deepcopy(ASSETS_DATASET_DEF)
db._add_var(v('ds_var1', 'EXAMPLE_LABEL.factor_1'))
db._add_var(v('ds_var2', 'ds_var1[t-1]', '0'))
sim = run_sim(
[
v('lagged_var',
'sum(EXAMPLE_LABEL.ds_var1[t-1] + EXAMPLE_LABEL.ds_var2)', '0')
],
policies=pols(attribute1=['1']),
dataset=db,
)
sim.assert_values_match(
'lagged_var',
[0, 48, 48, 48],
)
def test_multiclass_classifier():
vars = [
{
'short_name': 'prob',
'equation': 'CLF.predict_proba(t)'
},
]
# A dummy classifier with 3 classes. Not supported.
sim = run_sim(
[
v('prob', 'CLF.predict_proba(t)'),
],
models=clf_factory('dummy_multiclass.pickle'),
allow_errs=True,
)
sim.assert_errors_match([['CLF', '', 'binary']])
sim.assert_null_values('prob')
def test_find_threshold_iterative():
#Test the find_threshold_iterative() with
#multiple capacities.
data = setUp()
capacities = range(-50, 10**3, 50)
precision = .001
for capacity in capacities:
threshold = find_threshold_iterative(data, capacity, precision)
threshold = round_decimals_up(threshold, 3)
print(capacity, threshold)
if capacity < 0:
assert threshold == -1
continue
assert run_sim(data, threshold, capacity)
def test_sklearn_pipeline():
"""insurance_churn_model.pkl is an sklearn pipeline, with a classifier on
the tail end. It takes as input [number, number, number, number, string].
"""
ds = deepcopy(INSURANCE_DATASET_DEF)
ds._add_var(
v(
'churn',
'CLF.predict(t, t, initial_premium, initial_total_claims, location)'
))
sim = run_sim(
[],
models=clf_factory('insurance_churn_model.pkl'),
dataset=ds,
)
df = sim.dataset_df()
assert df.churn.isin([0, 1]).all()
def test_join_fn_reduced_dim_boolean_expr():
"""Verify that joins work even when the mask (the first arg) doesn't naturally
have the shape of a dataset variable.
"""
# This is another overmatching join, but it has the unusual feature that
# the first arg does not have dataset dimensionality. Want to make sure
# we just raise a user-visible join exception, rather than a hard crash.
sim = run_sim(
[
v('a_var', 't+4'),
v('b_var', "join(1==1, EXAMPLE_LABEL.factor_2)"),
],
dataset=ASSETS_DATASET_DEF,
num_sims=2,
num_steps=3,
allow_errs=True,
)
sim.assert_errors_match([['b_var', 'equation', 'Join']])
def test_sum_join_multiple_datasets_overmatching():
workers = deepcopy(WORKERS_DATASET_DEF)
workers.name = 'WORKERS'
more = deepcopy(MOREWORKERS_DATASET_DEF)
workers._add_var(
v('joined',
'sum(join(workers == MOREWORKERS.borkers, MOREWORKERS.widgets))'))
sim = run_sim(
[],
datasets=[more, workers],
num_sims=2,
)
# Second row of WORKERS matches two rows of MOREWORKERS having values of 20 and 2
sim.assert_values_match(
'WORKERS.joined',
[10, 20 + 2],
t=1,
)
def test_join_multiple_datasets_bijection():
"""A join involving two datasets which happen to have a perfect 1-to-1 correspondence.
"""
workers = deepcopy(WORKERS_DATASET_DEF)
workers.name = 'WORKERS'
para = PARAWORKERS_DATASET_DEF
workers._add_var(
v('joined', 'join(name == PARAWORKERS.name, PARAWORKERS.widgets)'))
sim = run_sim(
[],
datasets=[para, workers],
)
sim.assert_values_match(
'WORKERS.joined',
[10, 20],
t=1,
)
def test_external_model_basic():
# A regression model taking three inputs.
model = {
"label": "DEMAND_MODEL",
"localPath": "tests/sample_models/three_input_model.pkl",
}
sim = run_sim(
[
v('competitor_price', 'uniform(100, 200)'),
v('temperature', 'uniform(20, 30)'),
v('demand',
'DEMAND_MODEL.predict(competitor_price, temperature, t)'),
],
model=model,
)
df = sim.results
assert df.demand.isnull().sum() == 0
assert df.demand.std() > 0
def test_sum_join_multiple_datasets_bijection():
workers = deepcopy(WORKERS_DATASET_DEF)
workers.name = 'WORKERS'
para = deepcopy(PARAWORKERS_DATASET_DEF)
para.name = 'PARAWORKERS'
workers._add_var(
v('joined',
'sum(join(name == PARAWORKERS.name, PARAWORKERS.widgets))'))
sim = run_sim(
[],
datasets=[para, workers],
)
# Because this join is 1-1 the sum() should be a no-op
sim.assert_values_match(
'WORKERS.joined',
[10, 20],
t=1,
)
def test_join_multiple_datasets_injection():
"""Similar to above scenario, except the dataset we're joining to has an
extra row which doesn't participate in the join results
"""
workers = deepcopy(WORKERS_DATASET_DEF)
workers.name = 'WORKERS'
more = deepcopy(MOREWORKERS_DATASET_DEF)
workers._add_var(
v('joined', 'join(name == MOREWORKERS.name, MOREWORKERS.widgets)'))
sim = run_sim(
[],
datasets=[more, workers],
)
sim.assert_values_match(
'WORKERS.joined',
[10, 20],
t=1,
)
def test_with_dataset_formulas_accessing_raw_data_and_attributes_and_vars():
db = deepcopy(WORKERS_DATASET_DEF)
db._add_var(
v(
'rev_per_worker',
"EXAMPLE_LABEL.max_production / EXAMPLE_LABEL.workers * price_per_unit * a_var",
))
sim = run_sim(
[v('a_var', 't/10')],
policies=pols(price_per_unit=['10', '5']),
dataset=db,
num_sims=4,
)
sim.assert_values_match(
'EXAMPLE_LABEL.rev_per_worker',
[50, 10 / 3],
t=1,
policy=1,
)
def test_with_dataset_formulas_accessing_raw_data():
db = deepcopy(WORKERS_DATASET_DEF)
db._add_var(v('t_again', 't'))
sim = run_sim(
policies=pols(price_per_unit=['10', '1']),
dataset=db,
)
sim.assert_values_match(
'EXAMPLE_LABEL.t_again',
[
1,
1,
2,
2,
3,
3,
],
policy=0,
)
def test_probabilistic_sampling():
"""Ensure that .predict for a classifier samples rather than returning the argmax.
"""
vars = [
v('x', 'uniform(0, 1)'),
v('prob', 'CLF.predict_proba(x[t])'),
v('pred', 'CLF.predict(x[t])'),
]
sim = run_sim(
vars,
models=clf_factory('usually_positive_classifier.pickle'),
num_steps=1000,
)
# We trained a logistic classifier on noise in (0, 1), and with 90% positive
# labels. So all our predicted probabilities should be around .9, but we
# should still almost always get at least one negative prediction.
# (NB: There's a ~1e-46 of this test spuriously failing.)
df = sim.results
assert (df.prob > .5).all()
assert 0 in df.pred
assert 1 in df.pred
def test_sum_join_multiple_datasets_undermatching_multiple_policies():
"""NB: because dimensions of size 1 get a uniquely lax treatment under
numpy broadcasting rules, it's important to include some tests with >1
policy/sim, even if the quantity being tested has no dependence on policy
attributes or inter-simulation randomness.
"""
workers = deepcopy(WORKERS_DATASET_DEF)
workers.name = 'WORKERS'
more = deepcopy(MOREWORKERS_DATASET_DEF)
more._add_var(
v('joined', 'sum(join(name == WORKERS.name, WORKERS.max_production))'))
sim = run_sim(
[],
policies=DUMMY_POLS,
datasets=[more, workers],
num_sims=5,
)
# Second row of MOREWORKERS doesn't match any rows in WORKERS, hence sum is 0
sim.assert_values_match(
'MOREWORKERS.joined',
[200, 0, 100],
t=1,
policy=0,
)
def test_with_dataset_formulas_accessing_attributes():
db = deepcopy(WORKERS_DATASET_DEF)
db._add_var(
v(
'rev_per_worker',
"EXAMPLE_LABEL.max_production / EXAMPLE_LABEL.workers * price_per_unit"
))
sim = run_sim(
[],
policies=pols(price_per_unit=['10', '5']),
dataset=db,
num_sims=4,
)
sim.assert_values_match_across_time(
'EXAMPLE_LABEL.rev_per_worker',
[1000, 200 / 3],
policy=0,
)
sim.assert_values_match_across_time(
'EXAMPLE_LABEL.rev_per_worker',
[500, 100 / 3],
policy=1,
)
def test_arithmetic_index_expr():
sim = run_sim([
v('x', 't', '0'),
v('y', 'x[5-5]'),
])
sim.assert_values_match('y', [0, 0, 0, 0])
def test_constant_index_expr():
sim = run_sim([v('x', 't', '0'), v('y', 'x[0]')])
sim.assert_values_match('y', [0, 0, 0, 0])
def test_multiline_formula():
sim = run_sim([v('x', '1 +\n1')], )
sim.assert_values_match(
'x', 2) # broadcasting comparison - should be 2 everywhere
def test_fibonacci():
sim = run_sim([v('a', 'a[t-1] + a[t-2]', initial=1)], num_steps=5)
sim.assert_values_match('a', [1, 2, 3, 5, 8, 13])