def _get_RF(self, X_values, y_values, model_num, random_state):
rkf = RepeatedKFold(n_splits=5, n_repeats=4, random_state=random_state)
# RF
RF_model_errors = np.asarray([])
RF_resid = np.asarray([])
for train_index, test_index in rkf.split(X_values):
X_train, X_test = X_values[train_index], X_values[test_index]
y_train, y_test = y_values[train_index], y_values[test_index]
RF = rf.RF()
RF.train(X_train, y_train, model_num)
rf_pred, RF_errors = RF.predict(X_test, True)
rf_res = y_test - rf_pred
RF_model_errors = np.concatenate((RF_model_errors, RF_errors),
axis=None)
RF_resid = np.concatenate((RF_resid, rf_res), axis=None)
return RF_resid, RF_model_errors
def find_stats(X_values, y_values, stdev):
# define cross-validation splits
rkf = RepeatedKFold(n_splits=5, n_repeats=4, random_state=91936274)
# RF
print("finding rf scale factors")
RF_model_errors = np.asarray([])
RF_resid = np.asarray([])
for train_index, test_index in rkf.split(X_values):
#print("RF: {}".format(ctr))
#ctr = ctr + 1
X_train, X_test = X_values[train_index], X_values[test_index]
y_train, y_test = y_values[train_index], y_values[test_index]
RF = rf.RF()
RF.train_synth(X_train, y_train, std=stdev)
rf_pred, RF_errors = RF.predict_no_divide(X_test, True)
rf_res = y_test - rf_pred
RF_model_errors = np.concatenate((RF_model_errors, RF_errors),
axis=None)
RF_resid = np.concatenate((RF_resid, rf_res), axis=None)
abs_residuals = abs(RF_resid)
return abs_residuals, RF_model_errors
def _get_RF(self, X_train, y_train, X_test, y_test, model_num): RF = rf.RF() RF.train(X_train, y_train, model_num) predictions, model_errors = RF.predict(X_test, True) residuals = y_test - predictions return residuals, model_errors
y_train, y_test = Y.iloc[train_index_2], Y.iloc[test_index_2]
testGroup2 = np.delete(groups2, train_index_2)
if checkAlreadyDone(testGroup2[0], alreadyDone):
continue
frames = [X_test_1, X_test]
twoTest = pd.concat(frames)
yTest = [y_test_1, y_test]
yFrames = pd.concat(yTest)
testFinal = np.concatenate((testGroup, testGroup2))
RF = rf.RF()
RF.train(X_train, y_train, std=y_std)
GPR = gpr.GPR()
GPR.train(X_train,
y_train,
userkernel=gprsavedkernel,
std=y_std,
optimizer_restarts=0)
# Here instead of res, sigma try calculating domain prediction for the test data.
gpr_pred, GPR_errors = GPR.predict(twoTest, True)
rf_pred, RF_errors = RF.predict(twoTest, True)
RF_errors = rfslope * RF_errors + rfintercept
# Start measuring on different thresholds
def find_stats(X_values, y_values):
RF_model_errors = np.asarray([])
RF_residuals = np.asarray([])
GPR_model_errors = np.asarray([])
GPR_residuals = np.asarray([])
# define cross-validation splits
rkf = RepeatedKFold(n_splits=5, n_repeats=4, random_state=91936274)
#GPR
ctr = 1
for train_index, test_index in rkf.split(X_values):
print("GPR: {}/20 (iteration: {}/10)".format(ctr, outerctr))
ctr = ctr + 1
X_train, X_test = X_values[train_index], X_values[test_index]
y_train, y_test = y_values[train_index], y_values[test_index]
GPR = gpr.GPR()
GPR.train_synth(X_train,
y_train,
std=standard_deviation,
kernelchoice=1,
optimizer_restarts=10)
gpr_pred, gpr_errors = GPR.predict_no_divide(X_test, True)
gpr_res = (y_test - gpr_pred) / standard_deviation
gpr_errors = gpr_errors / standard_deviation
GPR_model_errors = np.concatenate((GPR_model_errors, gpr_errors),
axis=None)
GPR_residuals = np.concatenate((GPR_residuals, gpr_res), axis=None)
# define quantities to return
GPR_model_error_std = np.std(GPR_model_errors)
print("GPR standard deviation of model errors: {}".format(
GPR_model_error_std))
GPR_model_error_mean = np.mean(GPR_model_errors)
print("GPR mean of model errors: {}".format(GPR_model_error_mean))
cutoff = GPR_model_error_mean
# RF
print("finding rf scale factors")
RF_model_errors = np.asarray([])
RF_residuals = np.asarray([])
for train_index, test_index in rkf.split(X_values):
#print("RF: {}".format(ctr))
#ctr = ctr + 1
X_train, X_test = X_values[train_index], X_values[test_index]
y_train, y_test = y_values[train_index], y_values[test_index]
RF = rf.RF()
RF.train_synth(X_train, y_train, std=standard_deviation)
rf_pred, RF_errors = RF.predict_no_divide(X_test, True)
rf_res = y_test - rf_pred
RF_model_errors = np.concatenate((RF_model_errors, RF_errors),
axis=None)
RF_residuals = np.concatenate((RF_residuals, rf_res), axis=None)
abs_residuals = abs(RF_residuals)
res = np.asarray([])
sigma = np.asarray([])
# remove rf model errors and residuals that have gpr model error over cutoff
for i in range(0, len(GPR_model_errors)):
if GPR_model_errors[i] < cutoff:
res = np.append(res, abs_residuals[i])
sigma = np.append(sigma, RF_model_errors[i])
return cutoff, res, sigma