def run_xgb(args, steps_out):
#Parameter list:
param_list = ['speed', 'cos_wind_dir', 'sin_wind_dir']
predict = pd.DataFrame(columns={'speed', 'cos_wind_dir', 'sin_wind_dir'})
true = pd.DataFrame(columns={'speed', 'cos_wind_dir', 'sin_wind_dir'})
baseline = pd.DataFrame(columns={'speed', 'cos_wind_dir', 'sin_wind_dir'})
for param in param_list:
x_df, y_df, x, y = proc.prepare_x_y(measurement, forecast,
args.steps_in, steps_out, param)
X_train, X_test, y_train, y_test = train_test_split(x,
y,
test_size=0.2,
shuffle=False)
xg = XGBRegressor(max_depth=args.max_depth,
n_estimators=args.n_estimators,
colsample_bytree=args.colsample_bytree,
min_child_weight=args.min_child_weight,
subsample=args.subsample,
learning_rate=args.lr)
xg.fit(X_train, y_train)
y_hat = xg.predict(X_test)
predict[param] = pd.Series(y_hat)
#print(np.array(y_test).reshape(-1))
true[param] = pd.Series(
np.array(y_test).reshape(-1)) #y_test.flatten())
baseline[param] = x_df[param + '_forecast'][-len(y_hat):]
#reset index
baseline.reset_index(inplace=True)
return predict, true, baseline
def run_regression(steps_in, steps_out):
# Parameter list:
param_list = ['speed', 'cos_wind_dir', 'sin_wind_dir']
predict = pd.DataFrame(columns={'speed', 'cos_wind_dir', 'sin_wind_dir'})
true = pd.DataFrame(columns={'speed', 'cos_wind_dir', 'sin_wind_dir'})
baseline = pd.DataFrame(columns={'speed', 'cos_wind_dir', 'sin_wind_dir'})
for param in param_list:
x_df, y_df, x, y = proc.prepare_x_y(measurement, forecast, steps_in,
steps_out, param)
X_train, X_test, y_train, y_test = train_test_split(x,
y,
test_size=0.2,
shuffle=False)
xg = XGBRegressor(max_depth=5)
xg.fit(X_train, y_train)
y_hat = xg.predict(X_test)
predict[param] = pd.Series(y_hat)
true[param] = pd.Series(y_test.flatten())
baseline[param] = x_df[param + '_forecast'][-len(y_hat):]
# reset index
baseline.reset_index(inplace=True)
return predict, true, baseline
def train_xgb(measurement, forecast, steps_in, steps_out):
#flag message
print('running xgb for steps_out=', steps_out)
#Parameter list:
param_list = [
'scenario', 'dangerous'
] #['speed','cos_wind_dir','sin_wind_dir','scenario','dangerous']
for param in param_list:
print(param)
#train on the entire data
x_df, y_df, x, y = proc.prepare_x_y(measurement, forecast, steps_in,
steps_out, param)
#gridsearch
if param in ['speed', 'cos_wind_dir', 'sin_wind_dir']:
xgb_model = XGBRegressor()
splitter = KFold(n_splits=4, shuffle=True)
score = 'neg_mean_absolute_error'
if param in ['scenario', 'dangerous']:
xgb_model = XGBClassifier()
splitter = StratifiedKFold(n_splits=4, shuffle=True)
score = 'accuracy'
if (param == 'dangerous'):
sm = SMOTE(sampling_strategy=0.6, random_state=0)
x, y = sm.fit_resample(x, y)
score = 'roc_auc'
grid = GridSearchCV(xgb_model,
param_grid=grid_params,
scoring=score,
cv=splitter.split(x, y))
grid.fit(x, y)
best_model = grid.best_estimator_
#print grid parameters
print('gridsearch result for param: ', param)
print(grid.best_params_)
#save model into a pickle file
pickle.dump(
best_model,
open('trained_models/' + str(param) + '_t_' + str(steps_out),
'wb'))
return
def run_rf(steps_in, steps_out):
#flag message
print('running random forrest for steps_out=', steps_out)
#Parameter list:
param_list = [
'scenario', 'dangerous', 'speed', 'cos_wind_dir', 'sin_wind_dir'
] #['scenario','dangerous'] #
predict_test = pd.DataFrame(
columns={
'speed', 'cos_wind_dir', 'sin_wind_dir', 'scenario', 'dangerous',
'dangerous_proba'
})
predict_train = pd.DataFrame(
columns={
'speed', 'cos_wind_dir', 'sin_wind_dir', 'scenario', 'dangerous',
'dangerous_proba'
})
for param in param_list:
x_df, y_df, x, y = proc.prepare_x_y(measurement, forecast, steps_in,
steps_out, param)
x_train, x_test, y_train, y_test = train_test_split(x,
y,
test_size=0.2,
shuffle=False)
#gridsearch
if param in ['speed', 'cos_wind_dir', 'sin_wind_dir']:
rf_model = RandomForestRegressor()
splitter = KFold(n_splits=4, shuffle=True)
score = 'neg_mean_absolute_error' #MAE
if param in ['scenario', 'dangerous']:
rf_model = RandomForestClassifier()
splitter = StratifiedKFold(n_splits=4, shuffle=True)
score = 'accuracy'
# SMOTE for binary classification
if (param == 'dangerous'):
sm = SMOTE(sampling_strategy=0.6, random_state=0)
x_train, y_train = sm.fit_resample(x_train, y_train)
score = 'roc_auc'
grid = GridSearchCV(rf_model,
param_grid=grid_params,
scoring=score,
cv=splitter.split(x_train, y_train))
grid.fit(x_train, y_train)
print('gridsearch result for param: ', param)
print(grid.best_params_)
#save best parameters:
pickle.dump(
grid.best_params_,
open('results/params/rf_' + param + '_' + str(steps_out) + '.pkl',
'wb'))
best_model = grid.best_estimator_
#save model into a pickle file
pickle.dump(
best_model,
open(
'results/trained_models/rf_' + str(param) + '_' +
str(steps_out) + '.pkl', 'wb'))
#record results
predict_test[param] = pd.Series(best_model.predict(x_test))
predict_train[param] = pd.Series(best_model.predict(x_train))
if param == 'dangerous':
predict_test['dangerous_proba'] = pd.Series(
best_model.predict_proba(x_test)[:, 1])
predict_train['dangerous_proba'] = pd.Series(
best_model.predict_proba(x_train)[:, 1])
#record baseline and truth
predict_test['true'] = pd.Series(
np.array(y_test).reshape(-1)) #y_test.flatten())
predict_train['true'] = pd.Series(np.array(y_train).reshape(-1))
predict_test['baseline'] = x_df['dangerous_forecast'][-len(y_test
):]
predict_train['baseline'] = x_df['dangerous_forecast'][:len(y_train
)]
return predict_train, predict_test #, true, baseline