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
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.experimental import (  # noqa
    enable_hist_gradient_boosting)  # noqa
from sklearn.ensemble import (
    RandomForestClassifier, HistGradientBoostingClassifier)
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)
from mlprodict.onnxrt import OnnxInference


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["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, verbose=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)
            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={'': 14, 'ai.onnx.ml': 2},
                options=options)

        with open('model.onnx', 'wb') as f:
            f.write(model_onnx.SerializeToString())

        oinf = OnnxInference(model_onnx)
        inputs = {"input": df[[0, 1]].values.astype(numpy.float32),
                  "text": df[["text"]].values}
        pred_onx = oinf.run(inputs)

        diff = numpy.abs(
            pred_onx['probabilities'].ravel() -
            pipe.predict_proba(df).ravel()).sum()

        if verbose:
            def td(a):
                if hasattr(a, 'todense'):
                    b = a.todense()
                    ind = set(a.indices)
                    for i in range(b.shape[1]):
                        if i not in ind:
                            b[0, i] = numpy.nan
                    return b
                return a

            oinf = OnnxInference(model_onnx)
            pred_onx2 = oinf.run(inputs)
            diff2 = numpy.abs(
                pred_onx2['probabilities'].ravel() -
                pipe.predict_proba(df).ravel()).sum()

        if diff > 0.1:
            for i, (l1, l2) in enumerate(
                    zip(pipe.predict_proba(df),
                        pred_onx['probabilities'])):
                d = numpy.abs(l1 - l2).sum()
                if verbose and d > 0.1:
                    print("\nDISCREPENCY DETAILS")
                    print(d, i, l1, l2)
                    pre = pipe.steps[0][-1].transform(df)
                    print("idf", pre[i].dtype, td(pre[i]))
                    pre2 = pipe.steps[1][-1].transform(pre)
                    print("cas", pre2[i].dtype, td(pre2[i]))
                    inter = oinf.run(inputs, intermediate=True)
                    onx = inter['tfidftr_norm']
                    print("onx", onx.dtype, onx[i])
                    onx = inter['variable3']

        obs = dict(model=model.__name__,
                   discrepencies=diff,
                   model_onnx=model_onnx, pipe=pipe)
        if verbose:
            obs['discrepency2'] = diff2
        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

Out:

  0%|          | 0/4 [00:00<?, ?it/s]
 25%|##5       | 1/4 [00:00<00:00,  4.95it/s][02:54:54] WARNING: ../src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

 75%|#######5  | 3/4 [00:01<00:00,  2.81it/s]
100%|##########| 4/4 [00:01<00:00,  2.26it/s]
100%|##########| 4/4 [00:01<00:00,  2.46it/s]
                            model  discrepencies
0          RandomForestClassifier       0.000004
1  HistGradientBoostingClassifier            NaN
2                   XGBClassifier       9.675623
3                  LGBMClassifier       0.000008
model discrepencies error
0 RandomForestClassifier 0.000004 NaN
1 HistGradientBoostingClassifier NaN A sparse matrix was passed, but dense data is ...
2 XGBClassifier 9.675623 NaN
3 LGBMClassifier 0.000008 NaN


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

Out:

  0%|          | 0/4 [00:00<?, ?it/s]
 25%|##5       | 1/4 [00:00<00:00,  5.34it/s]
 50%|#####     | 2/4 [00:00<00:00,  2.54it/s][02:54:56] WARNING: ../src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

 75%|#######5  | 3/4 [00:02<00:01,  1.15s/it]
100%|##########| 4/4 [00:03<00:00,  1.11s/it]
100%|##########| 4/4 [00:03<00:00,  1.05it/s]
model discrepencies
0 RandomForestClassifier 0.000005
1 HistGradientBoostingClassifier 0.000007
2 XGBClassifier 0.000007
3 LGBMClassifier 0.000008


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])

Out:

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

Out:

  0%|          | 0/4 [00:00<?, ?it/s]
 25%|##5       | 1/4 [00:00<00:00,  4.37it/s][02:55:00] WARNING: ../src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

 75%|#######5  | 3/4 [00:02<00:00,  1.28it/s]
100%|##########| 4/4 [00:03<00:00,  1.21it/s]
100%|##########| 4/4 [00:03<00:00,  1.30it/s]
                            model  discrepencies
0          RandomForestClassifier      56.018917
1  HistGradientBoostingClassifier            NaN
2                   XGBClassifier       0.000007
3                  LGBMClassifier       0.000008
model discrepencies error
0 RandomForestClassifier 56.018917 NaN
1 HistGradientBoostingClassifier NaN A sparse matrix was passed, but dense data is ...
2 XGBClassifier 0.000007 NaN
3 LGBMClassifier 0.000008 NaN


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

Out:

  0%|          | 0/4 [00:00<?, ?it/s]
 25%|##5       | 1/4 [00:00<00:00,  4.29it/s][02:55:03] WARNING: ../src/learner.cc:1115: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

 75%|#######5  | 3/4 [00:01<00:00,  1.61it/s]
100%|##########| 4/4 [00:02<00:00,  1.53it/s]
100%|##########| 4/4 [00:02<00:00,  1.63it/s]
                            model  discrepencies
0          RandomForestClassifier      24.936082
1  HistGradientBoostingClassifier            NaN
2                   XGBClassifier       0.000007
3                  LGBMClassifier       0.000008
model discrepencies error
0 RandomForestClassifier 24.936082 NaN
1 HistGradientBoostingClassifier NaN A sparse matrix was passed, but dense data is ...
2 XGBClassifier 0.000007 NaN
3 LGBMClassifier 0.000008 NaN


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 11.202 seconds)

Gallery generated by Sphinx-Gallery