Note
Go to the end to download the full example code
TfIdf and sparse matrices#
TfidfVectorizer usually creates sparse data. If the data is sparse enough, matrices usually stays as sparse all along the pipeline until the predictor is trained. Sparse matrices do not consider null and missing values as they are not present in the datasets. Because some predictors do the difference, this ambiguity may introduces discrepencies when converter into ONNX. This example looks into several configurations.
Imports, setups#
All imports. It also registered onnx converters for :epgk:`xgboost` and lightgbm.
import warnings
import numpy
import pandas
import onnxruntime as rt
from tqdm import tqdm
from sklearn.compose import ColumnTransformer
from sklearn.datasets import load_iris
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.ensemble import RandomForestClassifier
try:
from sklearn.ensemble import HistGradientBoostingClassifier
except ImportError:
HistGradientBoostingClassifier = None
from xgboost import XGBClassifier
from lightgbm import LGBMClassifier
from skl2onnx.common.data_types import FloatTensorType, StringTensorType
from skl2onnx import to_onnx, update_registered_converter
from skl2onnx.sklapi import CastTransformer, ReplaceTransformer
from skl2onnx.common.shape_calculator import (
calculate_linear_classifier_output_shapes)
from onnxmltools.convert.xgboost.operator_converters.XGBoost import (
convert_xgboost)
from onnxmltools.convert.lightgbm.operator_converters.LightGbm import (
convert_lightgbm)
update_registered_converter(
XGBClassifier, 'XGBoostXGBClassifier',
calculate_linear_classifier_output_shapes, convert_xgboost,
options={'nocl': [True, False], 'zipmap': [True, False, 'columns']})
update_registered_converter(
LGBMClassifier, 'LightGbmLGBMClassifier',
calculate_linear_classifier_output_shapes, convert_lightgbm,
options={'nocl': [True, False], 'zipmap': [True, False]})
Artificial datasets#
Iris + a text column.
cst = ['class zero', 'class one', 'class two']
data = load_iris()
X = data.data[:, :2]
y = data.target
df = pandas.DataFrame(X)
df.columns = [f"c{c}" for c in df.columns]
df["text"] = [cst[i] for i in y]
ind = numpy.arange(X.shape[0])
numpy.random.shuffle(ind)
X = X[ind, :].copy()
y = y[ind].copy()
Train ensemble after sparse#
The example use the Iris datasets with artifical text datasets preprocessed with a tf-idf. sparse_threshold=1. avoids sparse matrices to be converted into dense matrices.
def make_pipelines(df_train, y_train, models=None,
sparse_threshold=1., replace_nan=False,
insert_replace=False):
if models is None:
models = [
RandomForestClassifier, HistGradientBoostingClassifier,
XGBClassifier, LGBMClassifier]
models = [_ for _ in models if _ is not None]
pipes = []
for model in tqdm(models):
if model == HistGradientBoostingClassifier:
kwargs = dict(max_iter=5)
elif model == XGBClassifier:
kwargs = dict(n_estimators=5, use_label_encoder=False)
else:
kwargs = dict(n_estimators=5)
if insert_replace:
pipe = Pipeline([
('union', ColumnTransformer([
('scale1', StandardScaler(), [0, 1]),
('subject',
Pipeline([
('count', CountVectorizer()),
('tfidf', TfidfTransformer()),
('repl', ReplaceTransformer()),
]), "text"),
], sparse_threshold=sparse_threshold)),
('cast', CastTransformer()),
('cls', model(max_depth=3, **kwargs)),
])
else:
pipe = Pipeline([
('union', ColumnTransformer([
('scale1', StandardScaler(), [0, 1]),
('subject',
Pipeline([
('count', CountVectorizer()),
('tfidf', TfidfTransformer())
]), "text"),
], sparse_threshold=sparse_threshold)),
('cast', CastTransformer()),
('cls', model(max_depth=3, **kwargs)),
])
try:
pipe.fit(df_train, y_train)
except TypeError as e:
obs = dict(model=model.__name__, pipe=pipe, error=e,
model_onnx=None)
pipes.append(obs)
continue
options = {model: {'zipmap': False}}
if replace_nan:
options[TfidfTransformer] = {'nan': True}
# convert
with warnings.catch_warnings(record=False):
warnings.simplefilter("ignore", (FutureWarning, UserWarning))
model_onnx = to_onnx(
pipe,
initial_types=[('input', FloatTensorType([None, 2])),
('text', StringTensorType([None, 1]))],
target_opset={'': 12, 'ai.onnx.ml': 2},
options=options)
with open('model.onnx', 'wb') as f:
f.write(model_onnx.SerializeToString())
sess = rt.InferenceSession(model_onnx.SerializeToString())
inputs = {"input": df[["c0", "c1"]].values.astype(numpy.float32),
"text": df[["text"]].values}
pred_onx = sess.run(None, inputs)
diff = numpy.abs(
pred_onx[1].ravel() -
pipe.predict_proba(df).ravel()).sum()
obs = dict(model=model.__name__,
discrepencies=diff,
model_onnx=model_onnx, pipe=pipe)
pipes.append(obs)
return pipes
data_sparse = make_pipelines(df, y)
stat = pandas.DataFrame(data_sparse).drop(['model_onnx', 'pipe'], axis=1)
if 'error' in stat.columns:
print(stat.drop('error', axis=1))
stat
0%| | 0/4 [00:00<?, ?it/s]
25%|##5 | 1/4 [00:00<00:00, 3.90it/s]
75%|#######5 | 3/4 [00:02<00:01, 1.03s/it]
100%|##########| 4/4 [00:03<00:00, 1.26it/s]
100%|##########| 4/4 [00:03<00:00, 1.25it/s]
model discrepencies
0 RandomForestClassifier 0.000005
1 HistGradientBoostingClassifier NaN
2 XGBClassifier 15.309917
3 LGBMClassifier 0.000009
Sparse data hurts.
Dense data#
Let’s replace sparse data with dense by using sparse_threshold=0.
data_dense = make_pipelines(df, y, sparse_threshold=0.)
stat = pandas.DataFrame(data_dense).drop(['model_onnx', 'pipe'], axis=1)
if 'error' in stat.columns:
print(stat.drop('error', axis=1))
stat
0%| | 0/4 [00:00<?, ?it/s]
25%|##5 | 1/4 [00:00<00:00, 4.38it/s]
50%|##### | 2/4 [00:05<00:06, 3.35s/it]
75%|#######5 | 3/4 [00:08<00:03, 3.12s/it]
100%|##########| 4/4 [00:09<00:00, 2.14s/it]
100%|##########| 4/4 [00:09<00:00, 2.31s/it]
This is much better. Let’s compare how the preprocessing applies on the data.
print("sparse")
print(data_sparse[-1]['pipe'].steps[0][-1].transform(df)[:2])
print()
print("dense")
print(data_dense[-1]['pipe'].steps[0][-1].transform(df)[:2])
sparse
(0, 0) -0.9006811702978088
(0, 1) 1.019004351971607
(0, 2) 0.4323732931220851
(0, 5) 0.9016947018779491
(1, 0) -1.1430169111851105
(1, 1) -0.13197947932162468
(1, 2) 0.4323732931220851
(1, 5) 0.9016947018779491
dense
[[-0.90068117 1.01900435 0.43237329 0. 0. 0.9016947 ]
[-1.14301691 -0.13197948 0.43237329 0. 0. 0.9016947 ]]
This shows RandomForestClassifier, XGBClassifier do not process the same way sparse and dense matrix as opposed to LGBMClassifier. And HistGradientBoostingClassifier fails.
Dense data with nan#
Let’s keep sparse data in the scikit-learn pipeline but replace null values by nan in the onnx graph.
data_dense = make_pipelines(df, y, sparse_threshold=1., replace_nan=True)
stat = pandas.DataFrame(data_dense).drop(['model_onnx', 'pipe'], axis=1)
if 'error' in stat.columns:
print(stat.drop('error', axis=1))
stat
0%| | 0/4 [00:00<?, ?it/s]
25%|##5 | 1/4 [00:00<00:00, 4.15it/s]
75%|#######5 | 3/4 [00:03<00:01, 1.16s/it]
100%|##########| 4/4 [00:05<00:00, 1.67s/it]
100%|##########| 4/4 [00:05<00:00, 1.46s/it]
model discrepencies
0 RandomForestClassifier 49.745050
1 HistGradientBoostingClassifier NaN
2 XGBClassifier 0.000004
3 LGBMClassifier 0.000009
Dense, 0 replaced by nan#
Instead of using a specific options to replace null values into nan values, a custom transformer called ReplaceTransformer is explicitely inserted into the pipeline. A new converter is added to the list of supported models. It is equivalent to the previous options except it is more explicit.
data_dense = make_pipelines(df, y, sparse_threshold=1., replace_nan=False,
insert_replace=True)
stat = pandas.DataFrame(data_dense).drop(['model_onnx', 'pipe'], axis=1)
if 'error' in stat.columns:
print(stat.drop('error', axis=1))
stat
0%| | 0/4 [00:00<?, ?it/s]
25%|##5 | 1/4 [00:00<00:00, 3.69it/s]
75%|#######5 | 3/4 [00:02<00:00, 1.23it/s]
100%|##########| 4/4 [00:02<00:00, 1.42it/s]
100%|##########| 4/4 [00:02<00:00, 1.44it/s]
model discrepencies
0 RandomForestClassifier 37.341543
1 HistGradientBoostingClassifier NaN
2 XGBClassifier 0.000004
3 LGBMClassifier 0.000009
Conclusion#
Unless dense arrays are used, because onnxruntime ONNX does not support sparse yet, the conversion needs to be tuned depending on the model which follows the TfIdf preprocessing.
Total running time of the script: ( 0 minutes 21.323 seconds)