def test_early_stopping(self):
np.random.seed(3187)
# Create a MERF model with a high early stopping threshold
m = MERF(max_iterations=10, gll_early_stop_threshold=0.1)
# Fit
m.fit(self.X_train, self.Z_train, self.clusters_train, self.y_train)
# The number of iterations should be less than max_iterations
self.assertTrue(len(m.gll_history) < 10)
def test_fit_and_predict_numpy(self):
m = MERF(max_iterations=5)
# Train
m.fit(np.array(self.X_train), np.array(self.Z_train), self.clusters_train, self.y_train)
# Predict Known Clusters
yhat_known = m.predict(np.array(self.X_known), np.array(self.Z_known), self.clusters_known)
self.assertEqual(len(yhat_known), 5)
# Predict New Clusters
yhat_new = m.predict(np.array(self.X_new), np.array(self.Z_new), self.clusters_new)
self.assertEqual(len(yhat_new), 2)
def test_fit_and_predict_pandas(self):
m = MERF(max_iterations=10)
# Train
m.fit(self.X_train, self.Z_train, self.clusters_train, self.y_train)
self.assertEqual(len(m.gll_history), 10)
# Predict Known Clusters
yhat_known = m.predict(self.X_known, self.Z_known, self.clusters_known)
self.assertEqual(len(yhat_known), 5)
# Predict New Clusters
yhat_new = m.predict(self.X_new, self.Z_new, self.clusters_new)
self.assertEqual(len(yhat_new), 2)
def test_user_defined_fe_model(self):
lgbm = LGBMRegressor()
m = MERF(fixed_effects_model=lgbm, max_iterations=5)
# Train
m.fit(self.X_train, self.Z_train, self.clusters_train, self.y_train)
self.assertEqual(len(m.gll_history), 5)
# Predict Known Clusters
yhat_known = m.predict(self.X_known, self.Z_known, self.clusters_known)
self.assertEqual(len(yhat_known), 5)
# Predict New Clusters
yhat_new = m.predict(self.X_new, self.Z_new, self.clusters_new)
self.assertEqual(len(yhat_new), 2)
def test_viz(self):
lgbm = LGBMRegressor()
m = MERF(fixed_effects_model=lgbm, max_iterations=5)
# Train
m.fit(
self.X_train,
self.Z_train,
self.clusters_train,
self.y_train,
self.X_known,
self.Z_known,
self.clusters_known,
self.y_known,
)
plot_merf_training_stats(m)
def test_validation_numpy(self):
m = MERF(max_iterations=3)
# Train
m.fit(
np.array(self.X_train),
np.array(self.Z_train),
self.clusters_train,
self.y_train,
np.array(self.X_new),
np.array(self.Z_new),
self.clusters_new,
self.y_new,
)
self.assertEqual(len(m.val_loss_history), 3)
# Predict Known Clusters
yhat_known = m.predict(self.X_known, self.Z_known, self.clusters_known)
self.assertEqual(len(yhat_known), 5)
# Predict New Clusters
yhat_new = m.predict(self.X_new, self.Z_new, self.clusters_new)
self.assertEqual(len(yhat_new), 2)
def test_pickle(self):
m = MERF(max_iterations=5)
# Train
m.fit(self.X_train, self.Z_train, self.clusters_train, self.y_train)
# Write to pickle file
with open("model.pkl", "wb") as fin:
pickle.dump(m, fin)
# Read back from pickle file
with open("model.pkl", "rb") as fout:
m_pkl = pickle.load(fout)
# Check that m is not the same object as m_pkl
self.assertIsNot(m_pkl, m)
# Predict Known Clusters
yhat_known_pkl = m_pkl.predict(self.X_known, self.Z_known, self.clusters_known)
yhat_known = m.predict(self.X_known, self.Z_known, self.clusters_known)
assert_almost_equal(yhat_known_pkl, yhat_known)
# Predict New Clusters
yhat_new_pkl = m_pkl.predict(self.X_new, self.Z_new, self.clusters_new)
yhat_new = m.predict(self.X_new, self.Z_new, self.clusters_new)
assert_almost_equal(yhat_new_pkl, yhat_new)
results.loc["Boosting_Ign","RMSE"] = np.sqrt(np.mean((y_test - y_pred) ** 2))
# 4. Gradient tree-boosting including the grouping variable as a categorical variable ('Boosting_Cat')
X_train_cat = np.column_stack((group_train,X_train))
X_test_cat = np.column_stack((group_test,X_test))
data_train_cat = gpb.Dataset(X_train_cat, y_train, categorical_feature=[0])
cvbst = gpb.cv(params=params, train_set=data_train_cat,
num_boost_round=1000, early_stopping_rounds=5,
nfold=4, verbose_eval=True, show_stdv=False, seed=1)
best_iter = np.argmin(cvbst['l2-mean'])
print("Best number of iterations: " + str(best_iter))
# Best number of iterations: 49
start_time = time.time() # measure time
bst = gpb.train(params=params, train_set=data_train_cat, num_boost_round=best_iter)
results.loc["Boosting_Cat","Time"] = time.time() - start_time
y_pred = bst.predict(data=X_test_cat)
results.loc["Boosting_Cat","RMSE"] = np.sqrt(np.mean((y_test - y_pred) ** 2))
# 5. Mixed-effects random forest ('MERF')
from merf import MERF
rf_params={'max_depth': 6, 'n_estimators': 300}
merf_model = MERF(max_iterations=100, rf_params=rf_params)
print("Warning: the following takes a lot of time")
start_time = time.time() # measure time
merf_model.fit(pd.DataFrame(X_train), np.ones(shape=(ntrain,1)), pd.Series(group_train), y_train)
results.loc["MERF","Time"] = time.time() - start_time
y_pred = merf_model.predict(pd.DataFrame(X_test), np.ones(shape=(ntrain,1)), pd.Series(group_test))
results.loc["MERF","RMSE"] = np.sqrt(np.mean((y_test - y_pred) ** 2))
print(results.apply(pd.to_numeric).round(3))
def mae(pred, true):
return np.mean(np.abs(pred - true))
# In[78]:
# Split data into train and test sets
indices = np.arange(len(y_samples))
X_train, X_test, y_train, y_test, idx_train, idx_test = train_test_split(
X_samples, y_samples, indices, test_size=0.2, random_state=301)
# In[ ]:
from merf import MERF
merf = MERF()
clusters_train = [ids[e] for e in idx_train]
clusters_train = pd.Series(clusters_train)
clusters_test = [ids[e] for e in idx_test]
clusters_test = pd.Series(clusters_test)
Z_train = np.ones(shape=(X_train.shape[0], 1))
merf.fit(X_train, Z_train, clusters_train, y_train)
train_preds = merf.predict(X_train, Z_train, clusters_train)
train_preds = np.maximum(train_preds, 0) # Don't predict negative cases
print('Train MAE:', mae(train_preds, y_train))
Z_test = np.ones(shape=(X_test.shape[0], 1))
test_preds = merf.predict(X_test, Z_test, clusters_test)
test_preds = np.maximum(test_preds, 0) # Don't predict negative cases
print('Test MAE:', mae(test_preds, y_test))
train = filter_train_by_visit(
visit_type, data['train'][data['train'].TRR_ID.isin(train_ids)])
if eq_train_ratio:
train = equalize_num_case_control(train, data['eq_cases_train_cols'])
if visit_type == "all":
if i > 0:
val_aurocs2 = []
for max_depth in [5, 10, 15]:
model = MERF(n_estimators=100,
gll_early_stop_threshold=0.001,
max_iterations=2,
max_depth=max_depth)
model.fit(train[cols],
pandas.DataFrame(np.ones((train.shape[0], 1))),
train.TRR_ID, train.diab_in_1_year)
test_y_hat = model.predict(
test[cols], pandas.DataFrame(np.ones((test.shape[0], 1))),
test.TRR_ID)
test_auroc = roc_auc_score(test.diab_in_1_year, test_y_hat)
val_aurocs2.append(test_auroc)
max_depth = [5, 10, 15][np.argmax(val_aurocs2)]
else:
max_depth = max_depths[np.argmax(val_aurocs)]
model = MERF(n_estimators=100,
gll_early_stop_threshold=0.001,
max_iterations=2,
max_depth=max_depth)
model.fit(train[cols], pandas.DataFrame(np.ones((train.shape[0], 1))),
def merf(normalise = False):
hyper_params = {
'task': 'train',
'boosting_type': 'gbdt',
'objective': 'regression',
'metric': ['l1', 'rmse'],
'learning_rate': 0.001,
'feature_fraction': 0.8,
"max_depth": 6,
"max_bin": 512,
"num_leaves": 40,
"num_iterations": 100000,
"n_estimators": 300,
"verbose": -1
}
# 'bagging_fraction': 0.7,
# 'bagging_freq': 10, "num_leaves": 12,
gbm = lgb.LGBMRegressor(**hyper_params)
ap2 = ap.fillna(method = "pad")
ap2.isna().sum().sum()
X_train, Y_train, X_test, Y_test = preprocessing(ap2, hour2int = True, onehotencode = False)
Z_train = np.ones((len(X_train), 1))
clusters_train = X_train['hours']
clusters_test= X_test['hours']
X_train1 = X_train.drop(["hours"],axis = 1)
X_test1 = X_test.drop(["hours"],axis = 1)
if normalise:
X_train1 =(X_train1-X_train1.mean())/X_train1.std()
X_test1 =(X_test1-X_test1.mean())/X_test1.std()
# we should not nornalise the Y (response)
# Y_train1 =(Y_train-Y_train.mean())/Y_train.std()
#my_imputer = SimpleImputer()
#X_train1 = my_imputer .fit_transform(X_train1) # fit missing
#X_test1 = my_imputer .fit_transform(X_test1)
# normalising for boosting is commonly not necessary, but for the mixed effect models
# we actually may want to normalise. But we only normalise X (predictors)!
# check if missing
print( Y_train1.isnull().any().any(),X_train1.isnull().any().any(),X_test.isnull().any().any())
merf = MERF(gbm, max_iterations = 4)
merf.fit(X_train1, Z_train, clusters_train, Y_train1)
Z_test = np.ones((len(X_test1), 1))
y_pred_ = merf.predict(X_test1, Z_test, clusters_test)
# also normalise the response and prediction wont work
#if normalise:
# y_pred = y_pred_*Y_train.std()+Y_train.mean()
mae = abs(y_pred - Y_test).mean()
rmse = math.sqrt(((y_pred - Y_test)*(y_pred - Y_test)).mean())
rrmse = rmse / Y_test.median()
r2 = get_r2_numpy_corrcoef(Y_test, y_pred)
return(mae, rmse, rrmse, r2)