def test_multiclass_confusion_matrix():
data = np.array([[0, 1], [0, 0], [1, 0], [1, 1], [1, 0], [1, 2], [2, 1],
[2, 0], [2, 2], [2, 1]])
pred = data.copy()
pred[3, 1] = 0
pred[4, 1] = 2
df = pd.DataFrame(data, columns=['feat', 'label'])
df2 = pd.DataFrame(pred, columns=['feat', 'label'])
favorable_values = [0, 1]
unfavorable_values = [2]
mcld = MulticlassLabelDataset(favorable_label=favorable_values,
unfavorable_label=unfavorable_values,
df=df,
label_names=['label'],
protected_attribute_names=['feat'])
mcld2 = MulticlassLabelDataset(favorable_label=favorable_values,
unfavorable_label=unfavorable_values,
df=df2,
label_names=['label'],
protected_attribute_names=['feat'])
cm = ClassificationMetric(mcld,
mcld2,
unprivileged_groups=[{
'feat': 2
}],
privileged_groups=[{
'feat': 0
}, {
'feat': 1
}])
confusion_matrix = cm.binary_confusion_matrix()
actual_labels_df = df[['label']].values
actual_labels_df2 = df2[['label']].values
assert np.all(actual_labels_df == mcld.labels)
assert np.all(actual_labels_df2 == mcld2.labels)
assert confusion_matrix == {'TP': 7.0, 'FN': 1.0, 'TN': 2.0, 'FP': 0.0}
fnr = cm.false_negative_rate_difference()
assert fnr == -0.2
def fairness_check(object_storage_url,
object_storage_username,
object_storage_password,
data_bucket_name,
result_bucket_name,
model_id,
feature_testset_path='processed_data/X_test.npy',
label_testset_path='processed_data/y_test.npy',
protected_label_testset_path='processed_data/p_test.npy',
model_class_file='model.py',
model_class_name='model',
favorable_label=0.0,
unfavorable_label=1.0,
privileged_groups=[{
'race': 0.0
}],
unprivileged_groups=[{
'race': 4.0
}]):
url = re.compile(r"https?://")
cos = Minio(url.sub('', object_storage_url),
access_key=object_storage_username,
secret_key=object_storage_password,
secure=False) # Local Minio server won't have HTTPS
dataset_filenamex = "X_test.npy"
dataset_filenamey = "y_test.npy"
dataset_filenamep = "p_test.npy"
weights_filename = "model.pt"
model_files = model_id + '/_submitted_code/model.zip'
cos.fget_object(data_bucket_name, feature_testset_path, dataset_filenamex)
cos.fget_object(data_bucket_name, label_testset_path, dataset_filenamey)
cos.fget_object(data_bucket_name, protected_label_testset_path,
dataset_filenamep)
cos.fget_object(result_bucket_name, model_id + '/' + weights_filename,
weights_filename)
cos.fget_object(result_bucket_name, model_files, 'model.zip')
# Load PyTorch model definition from the source code.
zip_ref = zipfile.ZipFile('model.zip', 'r')
zip_ref.extractall('model_files')
zip_ref.close()
modulename = 'model_files.' + model_class_file.split('.')[0].replace(
'-', '_')
'''
We required users to define where the model class is located or follow
some naming convention we have provided.
'''
model_class = getattr(importlib.import_module(modulename),
model_class_name)
# load & compile model
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model = model_class().to(device)
model.load_state_dict(torch.load(weights_filename, map_location=device))
"""Load the necessary labels and protected features for fairness check"""
x_test = np.load(dataset_filenamex)
y_test = np.load(dataset_filenamey)
p_test = np.load(dataset_filenamep)
_, y_pred = evaluate(model, x_test, y_test)
"""Calculate the fairness metrics"""
original_test_dataset = dataset_wrapper(
outcome=y_test,
protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
plain_predictions_test_dataset = dataset_wrapper(
outcome=y_pred,
protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
classified_metric_nodebiasing_test = ClassificationMetric(
original_test_dataset,
plain_predictions_test_dataset,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classified_metric_nodebiasing_test.true_positive_rate()
TNR = classified_metric_nodebiasing_test.true_negative_rate()
bal_acc_nodebiasing_test = 0.5 * (TPR + TNR)
print(
"#### Plain model - without debiasing - classification metrics on test set"
)
metrics = {
"Classification accuracy":
classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy":
bal_acc_nodebiasing_test,
"Statistical parity difference":
classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact":
classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference":
classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference":
classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index":
classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference":
classified_metric_nodebiasing_test.false_negative_rate_difference()
}
print("metrics: ", metrics)
return metrics
ClassificationMetric(
dataset_ground_truth,
dataset_classifier,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classificaltion_metric.true_positive_rate()
TNR = classificaltion_metric.true_negative_rate()
bal_acc_nodebiasing_test = 0.5 * (TPR + TNR)
metrics = {
"classification_accuracy":
classificaltion_metric.accuracy(),
"balanced_classification_accuracy":
bal_acc_nodebiasing_test,
"statistical_parity_difference":
classificaltion_metric.statistical_parity_difference(),
"disparate_impact":
classificaltion_metric.disparate_impact(),
"equal_opportunity_difference":
classificaltion_metric.equal_opportunity_difference(),
"average_odds_difference":
classificaltion_metric.average_odds_difference(),
"theil_index":
classificaltion_metric.theil_index(),
"false_negative_rate_difference":
classificaltion_metric.false_negative_rate_difference()
}
sys.stdout.write(json.dumps(metrics))
"#### Plain model - without debiasing - classification metrics on test set"
)
# print("Test set: Classification accuracy = %f" % classified_metric_nodebiasing_test.accuracy())
# print("Test set: Balanced classification accuracy = %f" % bal_acc_nodebiasing_test)
# print("Test set: Statistical parity difference = %f" % classified_metric_nodebiasing_test.statistical_parity_difference())
# print("Test set: Disparate impact = %f" % classified_metric_nodebiasing_test.disparate_impact())
# print("Test set: Equal opportunity difference = %f" % classified_metric_nodebiasing_test.equal_opportunity_difference())
# print("Test set: Average odds difference = %f" % classified_metric_nodebiasing_test.average_odds_difference())
# print("Test set: Theil index = %f" % classified_metric_nodebiasing_test.theil_index())
# print("Test set: False negative rate difference = %f" % classified_metric_nodebiasing_test.false_negative_rate_difference())
metrics = {
"Classification accuracy":
classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy":
bal_acc_nodebiasing_test,
"Statistical parity difference":
classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact":
classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference":
classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference":
classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index":
classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference":
classified_metric_nodebiasing_test.false_negative_rate_difference()
}
print("metrics: ", metrics)
def fairness_check(s3_url, bucket_name, s3_username, s3_password, training_id):
cos = boto3.resource("s3",
endpoint_url=s3_url,
aws_access_key_id=s3_username,
aws_secret_access_key=s3_password)
y_test_out = 'y_test.out'
p_test_out = 'p_test.out'
y_pred_out = 'y_pred.out'
get_s3_item(cos, bucket_name, training_id + '/' + y_test_out, y_test_out)
get_s3_item(cos, bucket_name, training_id + '/' + p_test_out, p_test_out)
get_s3_item(cos, bucket_name, training_id + '/' + y_pred_out, y_pred_out)
"""Need to generalize the protected features"""
unprivileged_groups = [{'race': 4.0}]
privileged_groups = [{'race': 0.0}]
favorable_label = 0.0
unfavorable_label = 1.0
"""Load the necessary labels and protected features for fairness check"""
y_test = np.loadtxt(y_test_out)
p_test = np.loadtxt(p_test_out)
y_pred = np.loadtxt(y_pred_out)
"""Calculate the fairness metrics"""
original_test_dataset = dataset_wrapper(outcome=y_test, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
plain_predictions_test_dataset = dataset_wrapper(outcome=y_pred, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
classified_metric_nodebiasing_test = ClassificationMetric(original_test_dataset,
plain_predictions_test_dataset,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classified_metric_nodebiasing_test.true_positive_rate()
TNR = classified_metric_nodebiasing_test.true_negative_rate()
bal_acc_nodebiasing_test = 0.5*(TPR+TNR)
print("#### Plain model - without debiasing - classification metrics on test set")
metrics = {
"Classification accuracy": classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy": bal_acc_nodebiasing_test,
"Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact": classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index": classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference()
}
print("metrics: ", metrics)
return metrics
def fairness_check(label_dir, model_dir):
"""Need to generalize the protected features"""
# races_to_consider = [0,4]
unprivileged_groups = [{'race': 4.0}]
privileged_groups = [{'race': 0.0}]
favorable_label = 0.0
unfavorable_label = 1.0
"""Load the necessary labels and protected features for fairness check"""
# y_train = np.loadtxt(label_dir + '/y_train.out')
# p_train = np.loadtxt(label_dir + '/p_train.out')
y_test = np.loadtxt(label_dir + '/y_test.out')
p_test = np.loadtxt(label_dir + '/p_test.out')
y_pred = np.loadtxt(label_dir + '/y_pred.out')
"""Calculate the fairness metrics"""
# original_traning_dataset = dataset_wrapper(outcome=y_train, protected=p_train,
# unprivileged_groups=unprivileged_groups,
# privileged_groups=privileged_groups,
# favorable_label=favorable_label,
# unfavorable_label=unfavorable_label)
original_test_dataset = dataset_wrapper(outcome=y_test, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
plain_predictions_test_dataset = dataset_wrapper(outcome=y_pred, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
classified_metric_nodebiasing_test = ClassificationMetric(original_test_dataset,
plain_predictions_test_dataset,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classified_metric_nodebiasing_test.true_positive_rate()
TNR = classified_metric_nodebiasing_test.true_negative_rate()
bal_acc_nodebiasing_test = 0.5*(TPR+TNR)
print("#### Plain model - without debiasing - classification metrics on test set")
# print("Test set: Classification accuracy = %f" % classified_metric_nodebiasing_test.accuracy())
# print("Test set: Balanced classification accuracy = %f" % bal_acc_nodebiasing_test)
# print("Test set: Statistical parity difference = %f" % classified_metric_nodebiasing_test.statistical_parity_difference())
# print("Test set: Disparate impact = %f" % classified_metric_nodebiasing_test.disparate_impact())
# print("Test set: Equal opportunity difference = %f" % classified_metric_nodebiasing_test.equal_opportunity_difference())
# print("Test set: Average odds difference = %f" % classified_metric_nodebiasing_test.average_odds_difference())
# print("Test set: Theil index = %f" % classified_metric_nodebiasing_test.theil_index())
# print("Test set: False negative rate difference = %f" % classified_metric_nodebiasing_test.false_negative_rate_difference())
metrics = {
"Classification accuracy": classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy": bal_acc_nodebiasing_test,
"Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact": classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index": classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference()
}
return {"metrics": metrics}
