def toy_train_val() -> TrainTestPair:
"""By making this a fixture, pytest can cache the result."""
data: DataTuple = em.toy().load()
train: DataTuple
test: DataTuple
train, test = em.train_test_split(data)
return TrainTestPair(train, test)
def test_empty_evaluate():
"""Test empty evaluate."""
empty_result = em.evaluate_models([em.toy()], repeats=3, delete_prev=True)
expected_result = pd.DataFrame(
[], columns=["dataset", "scaler", "transform", "model", "split_id"])
expected_result = expected_result.set_index(
["dataset", "scaler", "transform", "model", "split_id"])
pd.testing.assert_frame_equal(empty_result, expected_result)
def test_run_alg_suite():
"""Test run alg suite."""
dataset = em.adult(split="Race-Binary")
datasets: List[em.Dataset] = [dataset, em.toy()]
preprocess_models: List[em.PreAlgorithm] = [em.Upsampler()]
inprocess_models: List[em.InAlgorithm] = [em.LR(), em.SVM(kernel="linear")]
postprocess_models: List[em.PostAlgorithm] = []
metrics: List[em.Metric] = [em.Accuracy(), em.CV()]
per_sens_metrics: List[em.Metric] = [em.Accuracy(), em.TPR()]
parallel_results = em.evaluate_models_async(
datasets=datasets,
preprocess_models=preprocess_models,
inprocess_models=inprocess_models,
postprocess_models=postprocess_models,
metrics=metrics,
per_sens_metrics=per_sens_metrics,
repeats=1,
test_mode=True,
topic="pytest",
)
results = em.evaluate_models(
datasets,
preprocess_models,
inprocess_models,
postprocess_models,
metrics,
per_sens_metrics,
repeats=1,
test_mode=True,
delete_prev=True,
topic="pytest",
)
pd.testing.assert_frame_equal(parallel_results, results, check_like=True)
files = os.listdir(Path(".") / "results")
file_names = [
"pytest_Adult Race-Binary_Upsample uniform.csv",
"pytest_Adult Race-Binary_no_transform.csv",
"pytest_Toy_Upsample uniform.csv",
"pytest_Toy_no_transform.csv",
]
assert len(files) == 4
assert sorted(files) == file_names
for file in file_names:
written_file = pd.read_csv(Path(f"./results/{file}"))
assert (written_file["seed"][0], written_file["seed"][1]) == (0, 0)
assert written_file.shape == (2, 16)
reloaded = em.load_results("Adult Race-Binary", "Upsample uniform",
"pytest")
assert reloaded is not None
read = pd.read_csv(
Path(".") / "results" /
"pytest_Adult Race-Binary_Upsample uniform.csv")
read = read.set_index(
["dataset", "scaler", "transform", "model", "split_id"])
pd.testing.assert_frame_equal(reloaded, read)
def test_sequential_split():
"""Test sequential split."""
data: DataTuple = em.load_data(em.toy())
train: DataTuple
test: DataTuple
train, test, _ = em.SequentialSplit(train_percentage=0.8)(data)
assert all(data.x.iloc[0] == train.x.iloc[0])
assert all(data.x.iloc[-1] == test.x.iloc[-1])
assert len(train) == 320
assert len(test) == 80
def test_train_test_split():
"""Test train test split."""
data: DataTuple = em.load_data(em.toy())
train_test: Tuple[DataTuple, DataTuple] = em.train_test_split(data)
train, test = train_test
assert train is not None
assert test is not None
assert train.x.shape[0] > test.x.shape[0]
assert train.x["a1"].values[0] == 0.2365572108691669
assert train.x["a2"].values[0] == approx(0.008603090240657633, abs=1e-6)
assert train.x.shape[0] == train.s.shape[0]
assert train.s.shape[0] == train.y.shape[0]
num_samples = len(data)
len_default = math.floor((num_samples / 100) * 80)
assert train.s.shape[0] == len_default
assert test.s.shape[0] == num_samples - len_default
len_0_9 = math.floor((num_samples / 100) * 90)
train, test = em.train_test_split(data, train_percentage=0.9)
assert train.s.shape[0] == len_0_9
assert test.s.shape[0] == num_samples - len_0_9
len_0_7 = math.floor((num_samples / 100) * 70)
train, test = em.train_test_split(data, train_percentage=0.7)
assert train.s.shape[0] == len_0_7
assert test.s.shape[0] == num_samples - len_0_7
len_0_5 = math.floor((num_samples / 100) * 50)
train, test = em.train_test_split(data, train_percentage=0.5)
assert train.s.shape[0] == len_0_5
assert test.s.shape[0] == num_samples - len_0_5
len_0_3 = math.floor((num_samples / 100) * 30)
train, test = em.train_test_split(data, train_percentage=0.3)
assert train.s.shape[0] == len_0_3
assert test.s.shape[0] == num_samples - len_0_3
len_0_1 = math.floor((num_samples / 100) * 10)
train, test = em.train_test_split(data, train_percentage=0.1)
assert train.s.shape[0] == len_0_1
assert test.s.shape[0] == num_samples - len_0_1
len_0_0 = math.floor((num_samples / 100) * 0)
train, test = em.train_test_split(data, train_percentage=0.0)
assert train.s.shape[0] == len_0_0
assert train.name == "Toy - Train"
assert test.s.shape[0] == num_samples - len_0_0
assert test.name == "Toy - Test"
def test_random_seed():
"""Test random seed."""
data: DataTuple = em.load_data(em.toy())
train_test_0: Tuple[DataTuple, DataTuple] = em.train_test_split(data)
train_0, test_0 = train_test_0
assert train_0 is not None
assert test_0 is not None
assert train_0.x.shape[0] > test_0.x.shape[0]
assert train_0.x["a1"].values[0] == 0.2365572108691669
assert train_0.x["a2"].values[0] == approx(0.008603090240657633, abs=1e-6)
assert train_0.x.shape[0] == train_0.s.shape[0]
assert train_0.s.shape[0] == train_0.y.shape[0]
train_test_1: Tuple[DataTuple,
DataTuple] = em.train_test_split(data, random_seed=1)
train_1, test_1 = train_test_1
assert train_1 is not None
assert test_1 is not None
assert train_1.x.shape[0] > test_1.x.shape[0]
assert train_1.x["a1"].values[0] == 1.3736566330173798
assert train_1.x["a2"].values[0] == approx(0.21742296144957174, abs=1e-6)
assert train_1.x.shape[0] == train_1.s.shape[0]
assert train_1.s.shape[0] == train_1.y.shape[0]
train_test_2: Tuple[DataTuple,
DataTuple] = em.train_test_split(data, random_seed=2)
train_2, test_2 = train_test_2
assert train_2 is not None
assert test_2 is not None
assert train_2.x.shape[0] > test_2.x.shape[0]
assert train_2.x["a1"].values[0] == 1.2255705960148289
assert train_2.x["a2"].values[0] == -1.208089015454192
assert train_2.x.shape[0] == train_2.s.shape[0]
assert train_2.s.shape[0] == train_2.y.shape[0]
train_test_3: Tuple[DataTuple,
DataTuple] = em.train_test_split(data, random_seed=3)
train_3, test_3 = train_test_3
assert train_3 is not None
assert test_3 is not None
assert train_3.x.shape[0] > test_3.x.shape[0]
assert train_3.x["a1"].values[0] == approx(0.21165963748018515, abs=1e-6)
assert train_3.x["a2"].values[0] == -2.425137404779957
assert train_3.x.shape[0] == train_3.s.shape[0]
assert train_3.s.shape[0] == train_3.y.shape[0]
def test_threaded_agarwal():
"""Test threaded agarwal."""
models: List[InAlgorithmAsync] = [
Agarwal(dir='/tmp', classifier="SVM", fairness="EqOd")
]
class AssertResult(Metric):
_name = "assert_result"
def score(self, prediction, actual) -> float:
return (np.count_nonzero(prediction.hard.values == 1) == 45
and np.count_nonzero(prediction.hard.values == 0) == 35)
results = evaluate_models_async(datasets=[toy()],
inprocess_models=models,
metrics=[AssertResult()],
delete_prev=True)
assert results["assert_result"].iloc[0]
def test_run_alg_suite_wrong_metrics():
"""Test run alg suite wrong metrics."""
datasets: List[em.Dataset] = [em.toy(), em.adult()]
preprocess_models: List[em.PreAlgorithm] = [em.Upsampler()]
inprocess_models: List[em.InAlgorithm] = [em.SVM(kernel="linear"), em.LR()]
postprocess_models: List[em.PostAlgorithm] = []
metrics: List[em.Metric] = [em.Accuracy(), em.CV()]
per_sens_metrics: List[em.Metric] = [em.Accuracy(), em.TPR(), em.CV()]
with pytest.raises(em.MetricNotApplicable):
em.evaluate_models(
datasets,
preprocess_models,
inprocess_models,
postprocess_models,
metrics,
per_sens_metrics,
repeats=1,
test_mode=True,
)
def test_plot_evals():
"""Test plot evals."""
results: Results = evaluate_models(
datasets=[adult(), toy()],
preprocess_models=[Upsampler(strategy="preferential")],
inprocess_models=[LR(), SVM(kernel="linear"),
Kamiran()],
metrics=[Accuracy(), CV()],
per_sens_metrics=[TPR(), ProbPos()],
repeats=3,
test_mode=True,
delete_prev=True,
)
assert results["seed"][0] == results["seed"][1] == results["seed"][2] == 0
assert results["seed"][3] == results["seed"][4] == results["seed"][
5] == 2410
assert results["seed"][6] == results["seed"][7] == results["seed"][
8] == 4820
figs_and_plots: List[Tuple[plt.Figure,
plt.Axes]] # type: ignore[name-defined]
# plot with metrics
figs_and_plots = plot_results(results, Accuracy(), ProbPos())
# num(datasets) * num(preprocess) * num(accuracy combinations) * num(prop_pos combinations)
assert len(
figs_and_plots) == 2 * 2 * 1 * 2 + 4 # TODO: this +4 should be FIXED,
# it matches the column name containing a hyphen as a DIFF metric.
# plot with column names
figs_and_plots = plot_results(results, "Accuracy",
"prob_pos_sensitive-attr_0")
assert len(figs_and_plots) == 1 * 2 * 1 * 1
with pytest.raises(
ValueError,
match='No matching columns found for Metric "NMI preds and s".'):
plot_results(results, Accuracy(), NMI())
with pytest.raises(ValueError, match='No column named "unknown metric".'):
plot_results(results, "unknown metric", Accuracy())
def test_run_alg_suite_scaler():
"""Test run alg suite."""
dataset = em.adult(split="Race-Binary")
datasets: List[em.Dataset] = [dataset, em.toy()]
preprocess_models: List[em.PreAlgorithm] = [em.Upsampler()]
inprocess_models: List[em.InAlgorithm] = [em.LR(), em.SVM(kernel="linear")]
postprocess_models: List[em.PostAlgorithm] = []
metrics: List[em.Metric] = [em.Accuracy(), em.CV()]
per_sens_metrics: List[em.Metric] = [em.Accuracy(), em.TPR()]
results_no_scaler = em.evaluate_models(
datasets,
preprocess_models,
inprocess_models,
postprocess_models,
metrics,
per_sens_metrics,
repeats=1,
test_mode=True,
delete_prev=True,
topic="pytest",
)
results_scaler = em.evaluate_models(
datasets,
preprocess_models,
inprocess_models,
postprocess_models,
metrics,
per_sens_metrics,
scaler=StandardScaler(),
repeats=1,
test_mode=True,
delete_prev=True,
topic="pytest",
)
with pytest.raises(AssertionError):
pd.testing.assert_frame_equal(results_scaler,
results_no_scaler,
check_like=True)
def test_prop_train_test_split():
"""Test prop train test split."""
data: DataTuple = em.load_data(em.toy())
train: DataTuple
test: DataTuple
train, test, _ = ProportionalSplit(train_percentage=0.8)(data, split_id=0)
assert train is not None
assert test is not None
assert train.x.shape[0] > test.x.shape[0]
assert train.x["a1"].values[0] == -0.7135614558562237
assert train.x["a2"].values[0] == 1.1211390799513148
assert train.x.shape[0] == train.s.shape[0]
assert train.s.shape[0] == train.y.shape[0]
NUM_SAMPLES = len(data)
len_default = math.floor((NUM_SAMPLES / 100) * 80)
assert train.s.shape[0] == len_default
assert test.s.shape[0] == NUM_SAMPLES - len_default
# assert that the proportion of s=0 to s=1 has remained the same (and also test for y=0/y=1)
assert np.count_nonzero(train.s.to_numpy() == 0) == round(
0.8 * np.count_nonzero(data.s.to_numpy() == 0))
assert np.count_nonzero(train.y.to_numpy() == 0) == round(
0.8 * np.count_nonzero(data.y.to_numpy() == 0))
len_0_9 = math.floor((NUM_SAMPLES / 100) * 90)
train, test, _ = ProportionalSplit(train_percentage=0.9)(data, split_id=0)
assert train.s.shape[0] == len_0_9
assert test.s.shape[0] == NUM_SAMPLES - len_0_9
assert np.count_nonzero(train.s.to_numpy() == 0) == approx(round(
0.9 * np.count_nonzero(data.s.to_numpy() == 0)),
abs=1)
assert np.count_nonzero(train.y.to_numpy() == 0) == approx(round(
0.9 * np.count_nonzero(data.y.to_numpy() == 0)),
abs=1)
len_0_7 = math.floor((NUM_SAMPLES / 100) * 70)
train, test, _ = ProportionalSplit(train_percentage=0.7)(data, split_id=0)
assert train.s.shape[0] == len_0_7
assert test.s.shape[0] == NUM_SAMPLES - len_0_7
assert np.count_nonzero(train.s.to_numpy() == 0) == approx(round(
0.7 * np.count_nonzero(data.s.to_numpy() == 0)),
abs=1)
assert np.count_nonzero(train.y.to_numpy() == 0) == approx(round(
0.7 * np.count_nonzero(data.y.to_numpy() == 0)),
abs=1)
len_0_5 = math.floor((NUM_SAMPLES / 100) * 50)
train, test, _ = ProportionalSplit(train_percentage=0.5)(data, split_id=0)
assert train.s.shape[0] == len_0_5
assert test.s.shape[0] == NUM_SAMPLES - len_0_5
len_0_3 = math.floor((NUM_SAMPLES / 100) * 30)
train, test, _ = ProportionalSplit(train_percentage=0.3)(data, split_id=0)
assert train.s.shape[0] == len_0_3
assert test.s.shape[0] == NUM_SAMPLES - len_0_3
len_0_1 = math.floor((NUM_SAMPLES / 100) * 10)
train, test, _ = ProportionalSplit(train_percentage=0.1)(data, split_id=0)
assert train.s.shape[0] == len_0_1
assert test.s.shape[0] == NUM_SAMPLES - len_0_1
len_0_0 = math.floor((NUM_SAMPLES / 100) * 0)
train, test, _ = ProportionalSplit(train_percentage=0.0)(data, split_id=0)
assert train.s.shape[0] == len_0_0
assert train.name == "Toy - Train"
assert test.s.shape[0] == NUM_SAMPLES - len_0_0
assert test.name == "Toy - Test"
s = pd.DataFrame(np.random.randn(100), columns=["s"])
y = pd.DataFrame(np.random.randint(0, 5, 100), columns=["y"])
data = DataTuple(x=x, s=s, y=y)
train_test: Tuple[DataTuple, DataTuple] = train_test_split(data)
train, test = train_test
model: InAlgorithm = LR()
predictions: Prediction = model.run(train, test)
acc_per_sens = metric_per_sensitive_attribute(
predictions, test, TPR(pos_class=1,
labels=list(range(y.nunique()[0]))))
print(acc_per_sens)
PER_SENS = [
PerSensMetricTest(
dataset=toy(),
classifier=SVM(),
metric=Accuracy(),
expected_values={
"sensitive-attr_0": 0.921,
"sensitive-attr_1": 0.929
},
),
PerSensMetricTest(
dataset=toy(),
classifier=SVM(),
metric=ProbPos(),
expected_values={
"sensitive-attr_0": 0.368,
"sensitive-attr_1": 0.738
},
),
DT(
dataset=em.sqf(split="Race-Sex"),
samples=12_347,
x_features=144,
discrete_features=138,
s_features=1,
num_sens=4,
y_features=1,
num_labels=2,
name="SQF Race-Sex",
sum_s=24_336,
sum_y=1_289,
),
DT(
dataset=em.toy(),
samples=400,
x_features=10,
discrete_features=8,
s_features=1,
num_sens=2,
y_features=1,
num_labels=2,
name="Toy",
sum_s=200,
sum_y=231,
),
DT(
dataset=em.acs_income(
root=Path("~/Data"), year="2018", horizon=1, states=["AL"]),
samples=22_268,
def test_run_alg_suite_no_pipeline():
"""Run alg suite while avoiding the 'fair pipeline'."""
datasets: List[em.Dataset] = [em.toy(), em.adult()]
preprocess_models: List[em.PreAlgorithm] = [em.Upsampler()]
inprocess_models: List[em.InAlgorithm] = [
em.Kamiran(classifier="LR"), em.LR()
]
postprocess_models: List[em.PostAlgorithm] = []
metrics: List[em.Metric] = [em.Accuracy(), em.CV()]
per_sens_metrics: List[em.Metric] = [em.Accuracy(), em.TPR()]
parallel_results = em.evaluate_models_async(
datasets=datasets,
preprocess_models=preprocess_models,
inprocess_models=inprocess_models,
postprocess_models=postprocess_models,
metrics=metrics,
per_sens_metrics=per_sens_metrics,
repeats=1,
test_mode=True,
topic="pytest",
fair_pipeline=False,
)
results = em.evaluate_models(
datasets,
preprocess_models,
inprocess_models,
postprocess_models,
metrics,
per_sens_metrics,
repeats=1,
test_mode=True,
topic="pytest",
fair_pipeline=False,
delete_prev=True,
)
pd.testing.assert_frame_equal(parallel_results, results, check_like=True)
num_datasets = 2
num_preprocess = 1
num_fair_inprocess = 1
num_unfair_inprocess = 1
expected_num = num_datasets * (num_fair_inprocess +
(num_preprocess + 1) * num_unfair_inprocess)
assert len(results) == expected_num
kc_name = "Kamiran & Calders LR"
assert len(em.filter_results(results,
[kc_name])) == 2 # result for Toy and Adult
assert (len(em.filter_results(results, ["Toy"], index="dataset")) == 3
) # Kamiran, LR and Upsampler
different_name = em.filter_and_map_results(results,
{kc_name: "Kamiran & Calders"})
assert len(em.filter_results(different_name, [kc_name])) == 0
assert len(em.filter_results(different_name, ["Kamiran & Calders"])) == 2
pd.testing.assert_frame_equal(
em.filter_results(results, [kc_name]),
results.query(f"model == '{kc_name}'"),
)
