def _test_determenistic(distance, algorithm, weight, k): x_train, x_test, y_train, y_test = \ train_test_split(IRIS.data, IRIS.target, test_size=0.33, random_state=31) alg_results = [] for _ in range(N_TRIES): # models scikit_model = ScikitKNeighborsClassifier(n_neighbors=k, weights=weight, algorithm=algorithm, leaf_size=30, p=2, metric=distance) daal_model = DaalKNeighborsClassifier(n_neighbors=k, weights=weight, algorithm=algorithm, leaf_size=30, p=2, metric=distance) # training scikit_model.fit(x_train, y_train) daal_model.fit(x_train, y_train) # predict scikit_predict = scikit_model.predict(x_test) daal_predict = daal_model.predict(x_test) distances, indices = daal_model.kneighbors(x_test) alg_results.append((distances, indices, daal_predict)) # accuracy scikit_accuracy = accuracy_score(y_test, scikit_predict) daal_accuracy = accuracy_score(y_test, daal_predict) ratio = daal_accuracy / scikit_accuracy assert ratio >= ACCURACY_RATIO,\ f'kNN accuracy: scikit_accuracy={scikit_accuracy},daal_accuracy={daal_accuracy}, ratio={ratio}' # predict proba scikit_predict_proba = scikit_model.predict_proba(x_test) daal_predict_proba = daal_model.predict_proba(x_test) # log loss scikit_log_loss = log_loss(y_test, scikit_predict_proba) daal_log_loss = log_loss(y_test, daal_predict_proba) ratio = daal_log_loss / scikit_log_loss assert ratio <= LOG_LOSS_RATIO,\ f'kNN log_loss: scikit_log_loss={scikit_log_loss},daal_log_loss={daal_log_loss}, ratio={ratio}' # ROC AUC scikit_roc_auc = roc_auc_score(y_test, scikit_predict_proba, multi_class='ovr') daal_roc_auc = roc_auc_score(y_test, daal_predict_proba, multi_class='ovr') ratio = daal_roc_auc / scikit_roc_auc assert ratio >= ROC_AUC_RATIO,\ f'kNN roc_auc: scikit_roc_auc={scikit_roc_auc},daal_roc_auc={daal_roc_auc}, ratio={ratio}' for i in range(1, N_TRIES): for j, res in enumerate(alg_results[i]): assert (res == alg_results[0][j]).mean() == 1, \ f'Results are different between runs for {algorithm}, {weight}, {distance}, k={k}'
def check_determenistic(distance, algorithm, weight, k): x_train, x_test, y_train, y_test = make_dataset() alg_results = [] for _ in range(N_TRIES): alg = KNeighborsClassifier(n_neighbors=k, weights=weight, algorithm=algorithm, leaf_size=30, p=2, metric=distance) alg.fit(x_train, y_train) distances, indices = alg.kneighbors(x_test) labels = alg.predict(x_test) alg_results.append((distances, indices, labels)) accuracy = accuracy_score(labels, y_test) assert accuracy >= CHECK_RATIO_KNN,\ 'kNN classifier:accuracy={}'.format(accuracy) for i in range(1, N_TRIES): for j, res in enumerate(alg_results[i]): assert (res == alg_results[0][j]).mean() == 1, \ ('Results are different between runs for %s, %s, %s, k=%d'\ % (algorithm, weight, distance, k))