def test_lightgbm(tmp_path):
import lightgbm as lgb
X, y = simulate_data(500, 10, random_state=43210, classification=True)
train_data = lgb.Dataset(X, label=y)
param = {'objective': 'binary', 'metric': 'binary_logloss'}
num_round = 5
bst = lgb.train(param, train_data, num_round)
gbm_preds = bst.predict(X)
model_path = str(os.path.join(tmp_path, 'lgb.model'))
bst.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
fil_preds = np.asarray(fm.predict(X))
fil_preds = np.reshape(fil_preds, np.shape(gbm_preds))
assert array_equal(np.round(gbm_preds), fil_preds)
lcls = lgb.LGBMClassifier().set_params(objective='binary',
metric='binary_logloss')
lcls.fit(X, y)
gbm_proba = lcls.predict_proba(X)
lcls.booster_.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
fil_proba = np.asarray(fm.predict_proba(X))
fil_proba = np.reshape(fil_proba, np.shape(gbm_proba))
assert np.allclose(gbm_proba, fil_proba, 1e-2)
def _obtain_fil_model(treelite_handle,
depth,
output_class=True,
threshold=0.5,
algo='auto',
fil_sparse_format='auto'):
"""
Creates a Forest Inference (FIL) model using the treelite
handle obtained from the cuML Random Forest model.
Returns
----------
fil_model :
A Forest Inference model which can be used to perform
inferencing on the random forest model.
"""
storage_format = \
_check_fil_parameter_validity(depth=depth,
fil_sparse_format=fil_sparse_format,
algo=algo)
fil_model = ForestInference()
tl_to_fil_model = \
fil_model.load_from_randomforest(treelite_handle,
output_class=output_class,
threshold=threshold,
algo=algo,
storage_type=storage_format)
return tl_to_fil_model
def test_lightgbm(tmp_path, num_classes):
import lightgbm as lgb
X, y = simulate_data(500,
10 if num_classes == 2 else 50,
num_classes,
random_state=43210,
classification=True)
train_data = lgb.Dataset(X, label=y)
if num_classes == 2:
param = {
'objective': 'binary',
'metric': 'binary_logloss',
'num_class': 1
}
else:
param = {
'objective': 'ova', # 'multiclass', would use softmax
'metric': 'multi_logloss',
'num_class': num_classes
}
num_round = 5
model_path = str(os.path.join(tmp_path, 'lgb.model'))
bst = lgb.train(param, train_data, num_round)
bst.save_model(model_path)
if num_classes == 2:
# binary classification
gbm_proba = bst.predict(X)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
fil_proba = fm.predict_proba(X)
assert np.allclose(gbm_proba, fil_proba[:, 1], 1e-2)
gbm_preds = (gbm_proba > 0.5)
fil_preds = fm.predict(X)
assert array_equal(gbm_preds, fil_preds)
else:
# multi-class classification
# FIL doesn't yet support predict_proba() for multi-class
# TODO: Add a test for predict_proba() when it's supported
gbm_preds = bst.predict(X)
gbm_preds = gbm_preds.argmax(axis=1)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
fil_preds = fm.predict(X)
assert array_equal(np.round(gbm_preds), fil_preds)
def test_lightgbm(tmp_path, num_classes):
import lightgbm as lgb
X, y = simulate_data(500,
10 if num_classes == 2 else 50,
num_classes,
random_state=43210,
classification=True)
train_data = lgb.Dataset(X, label=y)
num_round = 5
model_path = str(os.path.join(tmp_path, 'lgb.model'))
if num_classes == 2:
param = {
'objective': 'binary',
'metric': 'binary_logloss',
'num_class': 1
}
bst = lgb.train(param, train_data, num_round)
bst.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
# binary classification
gbm_proba = bst.predict(X)
fil_proba = fm.predict_proba(X)[:, 1]
gbm_preds = (gbm_proba > 0.5)
fil_preds = fm.predict(X)
assert array_equal(gbm_preds, fil_preds)
np.testing.assert_allclose(gbm_proba,
fil_proba,
atol=proba_atol[num_classes > 2])
else:
# multi-class classification
lgm = lgb.LGBMClassifier(objective='multiclass',
boosting_type='gbdt',
n_estimators=num_round)
lgm.fit(X, y)
lgm.booster_.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
lgm_preds = lgm.predict(X)
assert array_equal(lgm.booster_.predict(X).argmax(axis=1), lgm_preds)
assert array_equal(lgm_preds, fm.predict(X))
# lightgbm uses float64 thresholds, while FIL uses float32
np.testing.assert_allclose(lgm.predict_proba(X),
fm.predict_proba(X),
atol=proba_atol[num_classes > 2])
def test_cpp_exception(tmp_path):
import lightgbm as lgb
num_class = 3
X, y = simulate_data(50,
5 * num_class,
num_class,
random_state=2020,
classification=True)
train_data = lgb.Dataset(X, label=y)
num_round = 1
param = {
'objective': 'ova', # 'multiclass', would use softmax
'metric': 'multi_logloss',
'num_class': num_class
}
bst = lgb.train(param, train_data, num_round)
model_path = str(os.path.join(tmp_path, 'lgb.model'))
bst.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type='lightgbm')
with pytest.raises(RuntimeError) as excinfo:
_ = fm.predict_proba(X)
assert ('predict_proba not supported for multi-class gradient boosted ' +
'decision trees' in str(excinfo.value))
def process(self, inputs):
input_meta = self.get_input_meta()
predict_col = self.conf.get('prediction', 'predict')
data_df = inputs[self.INPUT_PORT_NAME]
if self.INPUT_PORT_MODEL_NAME in input_meta:
# use external information instead of conf
filename = get_file_path(inputs[self.INPUT_PORT_MODEL_NAME])
train_cols = input_meta[self.INPUT_PORT_MODEL_NAME]['train']
train_cols = list(train_cols.keys())
else:
# use the conf information
filename = get_file_path(self.conf['file'])
if 'columns' in self.conf:
if self.conf.get('include', True):
train_cols = self.conf['columns']
else:
train_cols = [
col for col in data_df.columns
if col not in self.conf['columns']
]
# train_cols.sort()
fm = ForestInference.load(filename,
model_type=self.conf.get(
"model_type", "xgboost"))
prediction = fm.predict(data_df[train_cols])
prediction.index = data_df.index
data_df[predict_col] = prediction
return {self.OUTPUT_PORT_NAME: data_df}
def test_print_forest_shape(small_classifier_and_preds):
model_path, model_type, X, xgb_preds = small_classifier_and_preds
m = ForestInference.load(model_path, model_type=model_type,
output_class=True, compute_shape_str=True)
for substr in ['model size', ' MB', 'Depth histogram:', 'Leaf depth',
'Depth histogram fingerprint', 'Avg nodes per tree']:
assert substr in m.shape_str
def test_fil_skl_classification(n_rows, n_columns, n_estimators, max_depth,
storage_type, model_class):
# skip depth 20 for dense tests
if max_depth == 20 and not storage_type:
return
# settings
classification = True # change this to false to use regression
n_categories = 2
random_state = np.random.RandomState(43210)
X, y = simulate_data(n_rows,
n_columns,
n_categories,
random_state=random_state,
classification=classification)
# identify shape and indices
train_size = 0.80
X_train, X_validation, y_train, y_validation = train_test_split(
X, y, train_size=train_size, random_state=0)
init_kwargs = {
'n_estimators': n_estimators,
'max_depth': max_depth,
}
if model_class == RandomForestClassifier:
init_kwargs['max_features'] = 0.3
init_kwargs['n_jobs'] = -1
else:
# model_class == GradientBoostingClassifier
init_kwargs['init'] = 'zero'
skl_model = model_class(**init_kwargs)
skl_model.fit(X_train, y_train)
skl_preds = skl_model.predict(X_validation)
skl_preds_int = np.around(skl_preds)
skl_proba = skl_model.predict_proba(X_validation)
skl_acc = accuracy_score(y_validation, skl_preds > 0.5)
algo = 'NAIVE' if storage_type else 'BATCH_TREE_REORG'
fm = ForestInference.load_from_sklearn(skl_model,
algo=algo,
output_class=True,
threshold=0.50,
storage_type=storage_type)
fil_preds = np.asarray(fm.predict(X_validation))
fil_preds = np.reshape(fil_preds, np.shape(skl_preds_int))
fil_proba = np.asarray(fm.predict_proba(X_validation))
fil_proba = np.reshape(fil_proba, np.shape(skl_proba))
fil_acc = accuracy_score(y_validation, fil_preds)
assert fil_acc == pytest.approx(skl_acc, abs=1e-5)
assert array_equal(fil_preds, skl_preds_int)
assert np.allclose(fil_proba, skl_proba, 1e-3)
def test_lightgbm(tmp_path, num_classes):
import lightgbm as lgb
X, y = simulate_data(500,
10 if num_classes == 2 else 50,
num_classes,
random_state=43210,
classification=True)
train_data = lgb.Dataset(X, label=y)
if num_classes == 2:
param = {
'objective': 'binary',
'metric': 'binary_logloss',
'num_class': 1
}
else:
param = {
'objective': 'ova', # 'multiclass', would use softmax
'metric': 'multi_logloss',
'num_class': num_classes
}
num_round = 5
bst = lgb.train(param, train_data, num_round)
gbm_preds = bst.predict(X)
if num_classes > 2:
gbm_preds = gbm_preds.argmax(axis=1)
model_path = str(os.path.join(tmp_path, 'lgb.model'))
bst.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
fil_preds = fm.predict(X)
assert array_equal(np.round(gbm_preds), fil_preds)
if num_classes == 2:
lcls = lgb.LGBMClassifier().set_params(**param)
lcls.fit(X, y)
gbm_proba = lcls.predict_proba(X)
lcls.booster_.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
fil_proba = fm.predict_proba(X)
assert np.allclose(gbm_proba, fil_proba, 1e-2)
def test_fil_skl_classification(n_rows, n_columns, n_estimators, max_depth,
n_classes, storage_type, model_class):
# settings
classification = True # change this to false to use regression
random_state = np.random.RandomState(43210)
X, y = simulate_data(n_rows,
n_columns,
n_classes,
random_state=random_state,
classification=classification)
# identify shape and indices
train_size = 0.80
X_train, X_validation, y_train, y_validation = train_test_split(
X, y, train_size=train_size, random_state=0)
init_kwargs = {
'n_estimators': n_estimators,
'max_depth': max_depth,
}
if model_class == RandomForestClassifier:
init_kwargs['max_features'] = 0.3
init_kwargs['n_jobs'] = -1
else:
# model_class == GradientBoostingClassifier
init_kwargs['init'] = 'zero'
skl_model = model_class(**init_kwargs, random_state=random_state)
skl_model.fit(X_train, y_train)
skl_preds = skl_model.predict(X_validation)
skl_preds_int = np.around(skl_preds)
skl_proba = skl_model.predict_proba(X_validation)
skl_acc = accuracy_score(y_validation, skl_preds_int)
algo = 'NAIVE' if storage_type else 'BATCH_TREE_REORG'
fm = ForestInference.load_from_sklearn(skl_model,
algo=algo,
output_class=True,
threshold=0.50,
storage_type=storage_type)
fil_preds = np.asarray(fm.predict(X_validation))
fil_preds = np.reshape(fil_preds, np.shape(skl_preds_int))
fil_acc = accuracy_score(y_validation, fil_preds)
# fil_acc is within p99 error bars of skl_acc (diff == 0.017 +- 0.012)
# however, some tests have a delta as big as 0.04.
# sklearn uses float64 thresholds, while FIL uses float32
# TODO(levsnv): once FIL supports float64 accuracy, revisit thresholds
threshold = 1e-5 if n_classes == 2 else 0.1
assert fil_acc == pytest.approx(skl_acc, abs=threshold)
if n_classes == 2:
assert array_equal(fil_preds, skl_preds_int)
fil_proba = np.asarray(fm.predict_proba(X_validation))
fil_proba = np.reshape(fil_proba, np.shape(skl_proba))
assert np.allclose(fil_proba, skl_proba, 1e-3)
def test_output_args(small_classifier_and_preds):
model_path, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=False,
threshold=0.50)
X = np.asarray(X)
fil_preds = fm.predict(X)
assert np.allclose(fil_preds, xgb_preds, 1e-3)
def test_output_storage_type(storage_type, small_classifier_and_preds):
model_path, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
output_class=True,
storage_type=storage_type,
threshold=0.50)
xgb_preds_int = np.around(xgb_preds)
fil_preds = np.asarray(fm.predict(X))
assert np.allclose(fil_preds, xgb_preds_int, 1e-3)
def test_fil_skl_regression(n_rows, n_columns, n_estimators, max_depth,
storage_type, model_class):
# skip depth 20 for dense tests
if max_depth == 20 and storage_type == 'DENSE':
return
# settings
n_categories = 1
random_state = np.random.RandomState(43210)
X, y = simulate_data(n_rows,
n_columns,
n_categories,
random_state=random_state,
classification=False)
# identify shape and indices
train_size = 0.80
X_train, X_validation, y_train, y_validation = train_test_split(
X, y, train_size=train_size, random_state=0)
init_kwargs = {
'n_estimators': n_estimators,
'max_depth': max_depth,
}
if model_class == RandomForestRegressor:
init_kwargs['max_features'] = 0.3
init_kwargs['n_jobs'] = -1
else:
# model_class == GradientBoostingRegressor
init_kwargs['init'] = 'zero'
skl_model = model_class(**init_kwargs)
skl_model.fit(X_train, y_train)
skl_preds = skl_model.predict(X_validation)
skl_mse = mean_squared_error(y_validation, skl_preds)
algo = 'NAIVE' if storage_type == 'SPARSE' else 'BATCH_TREE_REORG'
fm = ForestInference.load_from_sklearn(skl_model,
algo=algo,
output_class=False,
storage_type=storage_type)
fil_preds = np.asarray(fm.predict(X_validation))
fil_preds = np.reshape(fil_preds, np.shape(skl_preds))
fil_mse = mean_squared_error(y_validation, fil_preds)
# if fil is better than skl, no need to fail the test
assert fil_mse <= skl_mse * (1. + 1e-7) + 1e-4
assert array_equal(fil_preds, skl_preds)
def test_thresholding(output_class, small_classifier_and_preds):
model_path, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=output_class,
threshold=0.50)
fil_preds = np.asarray(fm.predict(X))
if output_class:
assert ((fil_preds != 0.0) & (fil_preds != 1.0)).sum() == 0
else:
assert ((fil_preds != 0.0) & (fil_preds != 1.0)).sum() > 0
def test_output_algos(algo, small_classifier_and_preds):
model_path, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
algo=algo,
output_class=True,
threshold=0.50)
xgb_preds_int = np.around(xgb_preds)
fil_preds = np.asarray(fm.predict(X))
fil_preds = np.reshape(fil_preds, np.shape(xgb_preds_int))
assert np.allclose(fil_preds, xgb_preds_int, 1e-3)
def test_output_args(small_classifier_and_preds):
model_path, model_type, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
model_type=model_type,
algo='TREE_REORG',
output_class=False,
threshold=0.50)
X = np.asarray(X)
fil_preds = fm.predict(X)
fil_preds = np.reshape(fil_preds, np.shape(xgb_preds))
assert array_equal(fil_preds, xgb_preds, 1e-3)
def test_fil_classification(n_rows, n_columns, num_rounds, n_classes,
tmp_path):
# settings
classification = True # change this to false to use regression
random_state = np.random.RandomState(43210)
X, y = simulate_data(n_rows,
n_columns,
n_classes,
random_state=random_state,
classification=classification)
# identify shape and indices
n_rows, n_columns = X.shape
train_size = 0.80
X_train, X_validation, y_train, y_validation = train_test_split(
X, y, train_size=train_size, random_state=0)
model_path = os.path.join(tmp_path, 'xgb_class.model')
bst = _build_and_save_xgboost(model_path,
X_train,
y_train,
num_rounds=num_rounds,
classification=classification,
n_classes=n_classes)
dvalidation = xgb.DMatrix(X_validation, label=y_validation)
if n_classes == 2:
xgb_preds = bst.predict(dvalidation)
xgb_preds_int = np.around(xgb_preds)
xgb_proba = np.stack([1 - xgb_preds, xgb_preds], axis=1)
else:
xgb_proba = bst.predict(dvalidation)
xgb_preds_int = xgb_proba.argmax(axis=1)
xgb_acc = accuracy_score(y_validation, xgb_preds_int)
fm = ForestInference.load(model_path,
algo='auto',
output_class=True,
threshold=0.50)
fil_preds = np.asarray(fm.predict(X_validation))
fil_proba = np.asarray(fm.predict_proba(X_validation))
fil_acc = accuracy_score(y_validation, fil_preds)
assert fil_acc == pytest.approx(xgb_acc, abs=0.01)
assert array_equal(fil_preds, xgb_preds_int)
np.testing.assert_allclose(fil_proba,
xgb_proba,
atol=proba_atol[n_classes > 2])
def test_fil_classification(n_rows, n_columns, num_rounds, tmp_path):
# settings
classification = True # change this to false to use regression
n_rows = n_rows # we'll use 1 millions rows
n_columns = n_columns
n_categories = 2
random_state = np.random.RandomState(43210)
X, y = simulate_data(n_rows,
n_columns,
n_categories,
random_state=random_state,
classification=classification)
# identify shape and indices
n_rows, n_columns = X.shape
train_size = 0.80
X_train, X_validation, y_train, y_validation = train_test_split(
X, y, train_size=train_size, random_state=0)
model_path = os.path.join(tmp_path, 'xgb_class.model')
bst = _build_and_save_xgboost(model_path,
X_train,
y_train,
num_rounds=num_rounds,
classification=classification)
dvalidation = xgb.DMatrix(X_validation, label=y_validation)
xgb_preds = bst.predict(dvalidation)
xgb_preds_int = np.around(xgb_preds)
xgb_proba = np.stack([1 - xgb_preds, xgb_preds], axis=1)
xgb_acc = accuracy_score(y_validation, xgb_preds > 0.5)
fm = ForestInference.load(model_path,
algo='BATCH_TREE_REORG',
output_class=True,
threshold=0.50)
fil_preds = np.asarray(fm.predict(X_validation))
fil_preds = np.reshape(fil_preds, np.shape(xgb_preds_int))
fil_proba = np.asarray(fm.predict_proba(X_validation))
fil_proba = np.reshape(fil_proba, np.shape(xgb_proba))
fil_acc = accuracy_score(y_validation, fil_preds)
assert fil_acc == pytest.approx(xgb_acc, 0.01)
assert array_equal(fil_preds, xgb_preds_int)
assert array_equal(fil_proba, xgb_proba)
def test_output_blocks_per_sm(storage_type, blocks_per_sm,
small_classifier_and_preds):
model_path, model_type, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
model_type=model_type,
output_class=True,
storage_type=storage_type,
threshold=0.50,
blocks_per_sm=blocks_per_sm)
xgb_preds_int = np.around(xgb_preds)
fil_preds = np.asarray(fm.predict(X))
fil_preds = np.reshape(fil_preds, np.shape(xgb_preds_int))
assert np.allclose(fil_preds, xgb_preds_int, 1e-3)
def test_output_args(format, small_classifier_and_preds):
model_path, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=False,
threshold=0.50)
if format == 'numpy':
X = np.asarray(X)
elif format == 'cudf':
X = cudf.DataFrame.from_gpu_matrix(
cuda.to_device(np.ascontiguousarray(X)))
elif format == 'gpuarray':
X = cuda.to_device(np.ascontiguousarray(X))
else:
assert False
fil_preds = fm.predict(X)
assert np.allclose(fil_preds, xgb_preds, 1e-3)
def test_lightgbm(tmp_path):
import lightgbm as lgb
X, y = simulate_data(500, 10, random_state=43210, classification=True)
train_data = lgb.Dataset(X, label=y)
param = {'objective': 'binary', 'metric': 'binary_logloss'}
num_round = 5
bst = lgb.train(param, train_data, num_round)
gbm_preds = bst.predict(X)
model_path = str(os.path.join(tmp_path, 'lgb.model'))
bst.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=False,
model_type="lightgbm")
fil_preds = np.asarray(fm.predict(X))
assert np.allclose(gbm_preds, fil_preds, 1e-3)
def load_trained_model():
"""
Cached loading of trained [ XGBoost or RandomForest ] model into memory
Note: Models selected via filename parsing, edit if necessary
"""
xgb_models = glob.glob('/opt/ml/model/*_xgb')
rf_models = glob.glob('/opt/ml/model/*_rf')
kmeans_models = glob.glob('/opt/ml/model/*_kmeans')
app.logger.info(f'detected xgboost models : {xgb_models}')
app.logger.info(f'detected randomforest models : {rf_models}')
app.logger.info(f'detected kmeans models : {kmeans_models}\n\n')
model_type = None
start_time = time.perf_counter()
if len(xgb_models):
model_type = 'XGBoost'
model_filename = xgb_models[0]
if GPU_INFERENCE_FLAG:
# FIL
reloaded_model = ForestInference.load(model_filename)
else:
# native XGBoost
reloaded_model = xgboost.Booster()
reloaded_model.load_model(fname=model_filename)
elif len(rf_models):
model_type = 'RandomForest'
model_filename = rf_models[0]
reloaded_model = joblib.load(model_filename)
elif len(kmeans_models):
model_type = 'KMeans'
model_filename = kmeans_models[0]
reloaded_model = joblib.load(model_filename)
else:
raise Exception('! No trained models detected')
exec_time = time.perf_counter() - start_time
app.logger.info(f'> model {model_filename} '
f'loaded in {exec_time:.5f} s \n')
return reloaded_model, model_type, model_filename
def test_threads_per_tree(threads_per_tree,
small_classifier_and_preds):
model_path, model_type, X, xgb_preds = small_classifier_and_preds
fm = ForestInference.load(model_path,
model_type=model_type,
output_class=True,
storage_type='auto',
threshold=0.50,
threads_per_tree=threads_per_tree,
n_items=1)
fil_preds = np.asarray(fm.predict(X))
fil_proba = np.asarray(fm.predict_proba(X))
xgb_proba = np.stack([1-xgb_preds, xgb_preds], axis=1)
np.testing.assert_allclose(fil_proba, xgb_proba,
atol=proba_atol[False])
xgb_preds_int = np.around(xgb_preds)
fil_preds = np.reshape(fil_preds, np.shape(xgb_preds_int))
assert np.allclose(fil_preds, xgb_preds_int, 1e-3)
def test_fil_regression(n_rows, n_columns, num_rounds, tmp_path, max_depth):
# settings
classification = False # change this to false to use regression
n_rows = n_rows # we'll use 1 millions rows
n_columns = n_columns
random_state = np.random.RandomState(43210)
X, y = simulate_data(n_rows,
n_columns,
random_state=random_state,
classification=classification,
bias=10.0)
# identify shape and indices
n_rows, n_columns = X.shape
train_size = 0.80
X_train, X_validation, y_train, y_validation = train_test_split(
X, y, train_size=train_size, random_state=0)
model_path = os.path.join(tmp_path, 'xgb_reg.model')
bst = _build_and_save_xgboost(model_path,
X_train,
y_train,
classification=classification,
num_rounds=num_rounds,
xgboost_params={'max_depth': max_depth})
dvalidation = xgb.DMatrix(X_validation, label=y_validation)
xgb_preds = bst.predict(dvalidation)
xgb_mse = mean_squared_error(y_validation, xgb_preds)
fm = ForestInference.load(model_path,
algo='BATCH_TREE_REORG',
output_class=False)
fil_preds = np.asarray(fm.predict(X_validation))
fil_preds = np.reshape(fil_preds, np.shape(xgb_preds))
fil_mse = mean_squared_error(y_validation, fil_preds)
assert fil_mse == pytest.approx(xgb_mse, 0.01)
assert array_equal(fil_preds, xgb_preds)
def load_trained_model():
"""
Cached loading of trained [ XGBoost or RandomForest ] model into memory
Note: Models selected via filename parsing, edit if necessary
"""
xgb_models = glob.glob("/opt/ml/model/*_xgb")
rf_models = glob.glob("/opt/ml/model/*_rf")
app.logger.info(f"detected xgboost models : {xgb_models}")
app.logger.info(f"detected randomforest models : {rf_models}\n\n")
model_type = None
start_time = time.perf_counter()
if len(xgb_models):
model_type = "XGBoost"
model_filename = xgb_models[0]
if GPU_INFERENCE_FLAG:
# FIL
reloaded_model = ForestInference.load(model_filename)
else:
# native XGBoost
reloaded_model = xgboost.Booster()
reloaded_model.load_model(fname=model_filename)
elif len(rf_models):
model_type = "RandomForest"
model_filename = rf_models[0]
reloaded_model = joblib.load(model_filename)
else:
raise Exception("! No trained models detected")
exec_time = time.perf_counter() - start_time
app.logger.info(f"> model {model_filename} "
f"loaded in {exec_time:.5f} s \n")
return reloaded_model, model_type, model_filename
def test_lightgbm(tmp_path, num_classes, n_categorical):
import lightgbm as lgb
if n_categorical > 0:
n_features = 10
n_rows = 1000
n_informative = n_features
else:
n_features = 10 if num_classes == 2 else 50
n_rows = 500
n_informative = 'auto'
X, y = simulate_data(n_rows,
n_features,
num_classes,
n_informative=n_informative,
random_state=43210,
classification=True)
if n_categorical > 0:
X_fit, X_predict = to_categorical(X,
n_categorical=n_categorical,
invalid_frac=0.1,
random_state=43210)
else:
X_fit, X_predict = X, X
train_data = lgb.Dataset(X_fit, label=y)
num_round = 5
model_path = str(os.path.join(tmp_path, 'lgb.model'))
if num_classes == 2:
param = {'objective': 'binary',
'metric': 'binary_logloss',
'num_class': 1}
bst = lgb.train(param, train_data, num_round)
bst.save_model(model_path)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
# binary classification
gbm_proba = bst.predict(X_predict)
fil_proba = fm.predict_proba(X_predict)[:, 1]
gbm_preds = (gbm_proba > 0.5).astype(float)
fil_preds = fm.predict(X_predict)
assert array_equal(gbm_preds, fil_preds)
np.testing.assert_allclose(gbm_proba, fil_proba,
atol=proba_atol[num_classes > 2])
else:
# multi-class classification
lgm = lgb.LGBMClassifier(objective='multiclass',
boosting_type='gbdt',
n_estimators=num_round)
lgm.fit(X_fit, y)
lgm.booster_.save_model(model_path)
lgm_preds = lgm.predict(X_predict).astype(int)
fm = ForestInference.load(model_path,
algo='TREE_REORG',
output_class=True,
model_type="lightgbm")
assert array_equal(lgm.booster_.predict(X_predict).argmax(axis=1),
lgm_preds)
assert array_equal(lgm_preds, fm.predict(X_predict))
# lightgbm uses float64 thresholds, while FIL uses float32
np.testing.assert_allclose(lgm.predict_proba(X_predict),
fm.predict_proba(X_predict),
atol=proba_atol[num_classes > 2])
