def fit_evaluate(regr, X_train, X_val, y_train, y_val, log_y=False, scale=False, exclude_features=None):
print("Evaluating ...")
if y_val is None:
X_train, y_train = separate_X_y(X_train, exclude_features)
X_val, y_val = separate_X_y(X_val, exclude_features)
if scale:
scaler = RobustScaler()
scaler.fit(X_train)
X_train = scaler.transform(X_train)
# Fit on train, transforming the test, avoid data leak
X_val = scaler.transform(X_val)
if regr:
regr.verbose = False
if log_y:
regr.fit(X_train, np.log(y_train))
y_pred = np.exp(
np.array(regr.predict(X_val), dtype=np.float128))
else:
regr.fit(X_train, y_train)
y_pred = regr.predict(X_val)
else:
if log_y:
theta = normal_equation.normal_equation(
X_train, np.log(y_train))
y_pred = np.exp(customSGD.predict(theta, X_val))
else:
theta = normal_equation.normal_equation(X_train, y_train)
y_pred = customSGD.predict(theta, X_val)
evaluate(y_val, y_pred)
def kfold_evaluate(regr, folds, scoring, log_y=False, k=5):
rmse = []
mse = []
mae = []
r2 = []
i = 0
for fold in folds:
print("Evaluating %s" % (i))
(X_train, X_val, y_train, y_val) = fold
if regr == "customSGD":
if log_y:
theta = customSGD.SGD(lr=0.1, max_iter=20000,
X=X_train, y=np.log(y_train), lr_optimizer='invscaling',
print_interval=2000)
y_pred = np.exp(customSGD.predict(theta, X_val))
else:
theta = normal_equation.normal_equation(X_train, y_train)
y_pred = customSGD.predict(theta, X_val)
elif regr: # Any other Regressor from the SkLearn Library
regr.verbose = False
if log_y:
regr.fit(X_train, np.log(y_train))
y_pred = np.exp(
np.array(regr.predict(X_val), dtype=np.float128))
else:
regr.fit(X_train, y_train)
y_pred = regr.predict(X_val)
else:
if log_y:
theta = normal_equation.normal_equation(
X_train, np.log(y_train))
y_pred = np.exp(customSGD.predict(theta, X_val))
else:
theta = normal_equation.normal_equation(X_train, y_train)
y_pred = customSGD.predict(theta, X_val)
rmse.append(math.sqrt(((y_pred-y_val)**2).mean()))
mse.append(metrics.mean_squared_error(y_val, y_pred))
mae.append(metrics.mean_absolute_error(y_val, y_pred))
r2.append(metrics.r2_score(y_val, y_pred))
i += 1
print("RMSE: \t %.4f +/- %.4f" % (np.mean(rmse), np.std(rmse)))
print("MSE: \t %.4f +/- %.4f" % (np.mean(mse), np.std(mse)))
print("MAE: \t %.4f +/- %.4f" % (np.mean(mae), np.std(mae)))
print('R2: \t %.4f +/- %.4f' % (np.mean(r2), np.std(r2)))
def fit_eval_loss_customSGD(X_train, X_val, y_train, y_val, params={}, log_y=False, scale=False, exclude_features=None):
print("Evaluating ...")
if scale:
scaler = RobustScaler()
scaler.fit(X_train)
X_train = scaler.transform(X_train)
# Fit on train, transforming the test, avoid data leak
X_test = scaler.transform(X_val)
if log_y:
theta = customSGD.SGD(**params,X=X_train, y=np.log(y_train))
y_pred = np.exp(customSGD.predict(theta, X_test))
else:
theta = customSGD.SGD(**params,X=X_train, y=y_train)
y_pred = customSGD.predict(theta, X_test)
evaluate(y_val, y_pred)
def normal_equation_test():
X_, y_ = customSGD.get_toy_data_big()
X, X_val, y, y_val = model_selection.train_test_split(X_, y_, test_size=0.2, random_state=42)
theta = normal_equation(X, y)
y_pred = customSGD.predict(theta, X_val)
error = math.sqrt(((y_pred-y_val)**2).mean())
print("RMSE error: %.4f" % error)
print("MSE: %.3f" % metrics.mean_squared_error(y_val, y_pred))
print("MAE: %.3f" % metrics.mean_absolute_error(y_val, y_pred))
print('R2: %.3f' % metrics.r2_score(y_val, y_pred))