def helper(input_filename, expected_filename, config_file):
'''
'''
input_filename = os.path.join(BASE_DIR, input_filename)
expected_df = pd.read_csv(os.path.join(BASE_DIR, expected_filename))
if config_file:
config_file = os.path.join(BASE_DIR, config_file)
config = Configs.load_configs(config_file)
test_df, _ = audit(pd.read_csv(os.path.join(BASE_DIR, input_filename)),
config)
# match expected_df columns
shared_columns = [c for c in expected_df.columns if c in test_df.columns]
try:
expected_df = expected_df[shared_columns]
test_df = test_df[shared_columns]
combined_data = pd.merge(expected_df,
test_df,
on=['attribute_name', 'attribute_value'])
# subtract expected_df from test_df
except:
# collect output for
print('could not merge')
return (test_df, expected_df)
# see if close enough to 0
s = ""
EPS = 1e-6
for col in shared_columns:
if col not in {'attribute_value', 'attribute_name'}:
print('testing {} ...'.format(col))
try:
# TypeError: numpy boolean subtract, the `-` operator, is
# deprecated, use the bitwise_xor, the `^` operator, or the
# logical_xor function instead.
# found online that casting as float64 will go around, but
# would like to get Jesse's take on best way to avoid issue.
if np.mean(combined_data[col + "_x"].astype("float64") -
combined_data[col + "_y"].astype("float64")) > EPS:
exp_mean = np.mean(combined_data[col + "_x"])
aeq_mean = np.mean(combined_data[col + "_y"])
s += "{} fails: Expected {} on average, but aequitas returned {}\n".format(
col, exp_mean, aeq_mean)
pytest.fail(s)
except:
if not all(combined_data[col + "_x"] == combined_data[col +
"_y"]):
s += "{} fails: at least one entry was not the same between data sets\n".format(
col)
pytest.fail(s)
def helper(input_filename, expected_filename, config_file):
'''
'''
input_filename = os.path.join(BASE_DIR, input_filename)
expected_df = pd.read_csv(os.path.join(BASE_DIR, expected_filename))
if config_file:
config_file = os.path.join(BASE_DIR, config_file)
config = Configs.load_configs(config_file)
test_df, _ = audit(pd.read_csv(os.path.join(BASE_DIR, input_filename)),
config)
# match expected_df columns
shared_columns = [c for c in expected_df.columns if c in test_df.columns]
try:
expected_df = expected_df[shared_columns]
test_df = test_df[shared_columns]
combined_data = pd.merge(expected_df,
test_df,
on=['attribute_name', 'attribute_value'])
# subtract expected_df from test_df
except:
# collect output for
print('could not merge')
return (test_df, expected_df)
# see if close enough to 0
s = ""
EPS = 1e-6
for col in shared_columns:
if col not in {'attribute_value', 'attribute_name'}:
print('testing {} ...'.format(col))
try:
if np.mean(combined_data[col + "_x"] -
combined_data[col + "_y"]) > EPS:
exp_mean = np.mean(combined_data[col + "_x"])
aeq_mean = np.mean(combined_data[col + "_y"])
s += "{} fails: Expected {} on average, but aequitas returned {}\n".format(
col, exp_mean, aeq_mean)
pytest.fail(s)
except:
if not all(combined_data[col + "_x"] == combined_data[col +
"_y"]):
s += "{} fails: at least one entry was not the same between data sets\n".format(
col)
pytest.fail(s)
def audit_file(name, dirname):
upload_path = os.path.join(tempfile.gettempdir(), dirname)
data_path = os.path.join(upload_path, name + '.csv')
if not os.path.exists(data_path):
abort(404)
try:
df = pd.read_csv(data_path)
except pd.errors.ParserError:
flash('Bad CSV file – could not parse', 'warning')
return redirect(url_for('home'))
(df, groups) = preprocess_input_df(df)
if "submit" not in request.form:
subgroups = {col: list(set(df[col])) for col in groups}
# set defaults
for (key, values) in (
('race', ('White', 'Caucasian')),
('sex', ('Male',)),
('gender', ('Male',)),
('age_cat', ('25 - 45',)),
('education', ('HS-grad',)),
):
if key in subgroups:
subgroups[key].sort(key=lambda value: int(value not in values))
supported_fairness_measures = Fairness().get_fairness_measures_supported(df)
fairness_measures = [x for x in FAIR_MAP_ORDER if FAIR_MAP[x].issubset(set(supported_fairness_measures))]
return render_template('audit.html',
categories=groups,
subcategories=subgroups,
fairness=fairness_measures)
rgm = request.form["ref_groups_method"]
if rgm == 'predefined':
group_variables = request.form.getlist('group_variable1')
else:
group_variables = request.form.getlist('group_variable2')
# check if user forgot to select anything; return all
if len(group_variables) == 0:
group_variables = groups
# remove unwanted cols from df
subgroups = {g: request.form[g] for g in group_variables}
# majority_groups = request.form.getlist('use_majority_group')
raw_fairness_measures = request.form.getlist('fairness_measures')
if len(raw_fairness_measures) == 0:
fairness_measures = list(Fairness().get_fairness_measures_supported(df))
else:
# map selected measures to input
fairness_measures = [y for x in raw_fairness_measures for y in FAIR_MAP[x]]
try:
fv = float(request.form['fairness_pct'])
except (KeyError, ValueError):
fv = None
fp = fv / 100.0 if fv else 0.8
configs = Configs(ref_groups=subgroups,
ref_groups_method=rgm,
fairness_threshold=fp,
fairness_measures=fairness_measures,
attr_cols=group_variables)
(_gv_df, report) = audit(df,
# model_id=1,
configs=configs,
preprocessed=True)
for reportid in itertools.count(1):
report_path = os.path.join(upload_path, str(reportid))
if not os.path.exists(report_path):
break
with open(report_path, 'w') as fd:
fd.write(report)
return redirect(url_for("report",
dirname=dirname,
name=name,
reportid=reportid))