def model_xgboost(data):
model = XGBClassifier()
model.fit(data['X_train'], data['y_train'])
y_valid_hat = model.predict(data['X_valid'])
BER = src.ber(y=data['y_valid'].tolist(), y_hat=y_valid_hat.tolist())
return BER
def do_predict(idx, params):
# do prediction
# load model
model_path = os.path.join('cache', 'models_fitted',
(params['hash_id'] + '.pkl'))
if os.path.isfile(model_path):
model_fitted = pickle.load(open(model_path, 'rb'))
# load data
data_path = os.path.join('cache', 'data_reduced',
(params['data_hash_id'] + '.pkl'))
data = pickle.load(open(data_path, 'rb'))
# predict validation set
y_valid_hat = model_fitted.predict(data['X_valid'])
# output BER
ber = src.ber(y=data['y_valid'].tolist(), y_hat=y_valid_hat.tolist())
roc_auc = roc_auc_score(y_true=data['y_valid'].tolist(),
y_score=y_valid_hat.tolist())
f1 = f1_score(y_true=data['y_valid'].tolist(),
y_pred=y_valid_hat.tolist())
predict_ber = {
'hash_id': params['hash_id'],
'BER': ber,
'ROC_AUC': roc_auc,
'F1': f1
}
predict_ber = pd.DataFrame(predict_ber, index=[idx])
predict_ber.to_csv(
os.path.join('cache', 'predict_ber', (params['hash_id'] + '.csv')))
print("Predicted BER: " + str(ber))
def model_svm(data):
X = data['X_train']
y = data['y_train']
print(X.shape)
X_new = SelectPercentile(mutual_info_classif,
percentile=10).fit_transform(X, y)
print(X_new.shape)
c = [2**float(ci) for ci in range(-5, 11, 2)]
g = [2**float(gi) for gi in range(-11, 9, 2)]
degree = [d for d in range(0, 6)]
kernel = ['poly'] #, 'sigmoid', 'rbf']
params = {'c': c, 'g': g, 'degree': degree, 'kernel': kernel}
grid = src.expand_grid(**params)
grid = pd.DataFrame(grid)
grid['BERcv'] = 1.0
grid.shape
if False:
from sklearn.utils import shuffle
grid = shuffle(grid)
for idx, row in grid[0:20].iterrows():
model = svm.SVC(gamma=row['g'],
C=row['c'],
kernel=row['kernel'],
degree=row['degree'])
BERs = []
kf = KFold(n_splits=10, random_state=1986, shuffle=True)
for train_index, test_index in kf.split(X):
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
model.fit(X_train, y_train)
y_test_hat = model.predict(X_test)
BERs.append(src.ber(y=y_test.tolist(), y_hat=y_test_hat.tolist()))
grid.loc[idx, 'BERcv'] = np.mean(BERs)
best = grid[grid['BERcv'] == grid['BERcv'].min()]
model = svm.SVC(gamma=np.median(best['g']),
C=np.median(best['c']),
kernel=row['kernel'])
return BER
def benchmark_pca_logistic(data):
# make PCA Logistic reg
pca = PCA(n_components=5).fit(data['X_train'])
Xpca_train = pca.transform(data['X_train'])
Xpca_valid = pca.transform(data['X_valid'])
# Basic Logistic regression
clf = LogisticRegression(random_state=0,
solver='sag',
multi_class='ovr',
max_iter=1000).fit(Xpca_train, data['y_train'])
y_valid_hat = clf.predict(Xpca_valid)
BER = src.ber(y=data['y_valid'].tolist(), y_hat=y_valid_hat.tolist())
return BER
def model_lightgbm(data):
model = lgbm.LGBMClassifier()
model.fit(data['X_train'], data['y_train'])
y_valid_hat = model.predict(data['X_valid'])
BER = src.ber(y=data['y_valid'].tolist(), y_hat=y_valid_hat.tolist())
return BER
def benchmark_all_neg1(data):
y_valid_hat = [-1 for x in range(len(data['y_valid'].tolist()))]
BER = src.ber(y=data['y_valid'].tolist(), y_hat=y_valid_hat)
return BER