TfIdfVectorizer with ONNX

This example is inspired from the following example: Column Transformer with Heterogeneous Data Sources which builds a pipeline to classify text.

Train a pipeline with TfidfVectorizer

It replicates the same pipeline taken from scikit-learn documentation but reduces it to the part ONNX actually supports without implementing a custom converter. Let’s get the data.

import matplotlib.pyplot as plt
import os
from onnx.tools.net_drawer import GetPydotGraph, GetOpNodeProducer
import numpy
import onnxruntime as rt
from skl2onnx.common.data_types import StringTensorType
from skl2onnx import convert_sklearn
import numpy as np

from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.datasets import fetch_20newsgroups
try:
    from sklearn.datasets._twenty_newsgroups import (
        strip_newsgroup_footer, strip_newsgroup_quoting)
except ImportError:
    # scikit-learn < 0.24
    from sklearn.datasets.twenty_newsgroups import (
        strip_newsgroup_footer, strip_newsgroup_quoting)
from sklearn.decomposition import TruncatedSVD
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.metrics import classification_report
from sklearn.linear_model import LogisticRegression


# limit the list of categories to make running this example faster.
categories = ['alt.atheism', 'talk.religion.misc']
train = fetch_20newsgroups(random_state=1,
                           subset='train',
                           categories=categories,
                           )
test = fetch_20newsgroups(random_state=1,
                          subset='test',
                          categories=categories,
                          )

The first transform extract two fields from the data. We take it out form the pipeline and assume the data is defined by two text columns.

class SubjectBodyExtractor(BaseEstimator, TransformerMixin):
    """Extract the subject & body from a usenet post in a single pass.
    Takes a sequence of strings and produces a dict of sequences. Keys are
    `subject` and `body`.
    """

    def fit(self, x, y=None):
        return self

    def transform(self, posts):
        # construct object dtype array with two columns
        # first column = 'subject' and second column = 'body'
        features = np.empty(shape=(len(posts), 2), dtype=object)
        for i, text in enumerate(posts):
            headers, _, bod = text.partition('\n\n')
            bod = strip_newsgroup_footer(bod)
            bod = strip_newsgroup_quoting(bod)
            features[i, 1] = bod

            prefix = 'Subject:'
            sub = ''
            for line in headers.split('\n'):
                if line.startswith(prefix):
                    sub = line[len(prefix):]
                    break
            features[i, 0] = sub

        return features


train_data = SubjectBodyExtractor().fit_transform(train.data)
test_data = SubjectBodyExtractor().fit_transform(test.data)

The pipeline is almost the same except we remove the custom features.

pipeline = Pipeline([
    ('union', ColumnTransformer(
        [
            ('subject', TfidfVectorizer(min_df=50, max_features=500), 0),

            ('body_bow', Pipeline([
                ('tfidf', TfidfVectorizer()),
                ('best', TruncatedSVD(n_components=50)),
            ]), 1),

            # Removed from the original example as
            # it requires a custom converter.
            # ('body_stats', Pipeline([
            #   ('stats', TextStats()),  # returns a list of dicts
            #   ('vect', DictVectorizer()),  # list of dicts -> feature matrix
            # ]), 1),
        ],

        transformer_weights={
            'subject': 0.8,
            'body_bow': 0.5,
            # 'body_stats': 1.0,
        }
    )),

    # Use a LogisticRegression classifier on the combined features.
    # Instead of LinearSVC (not fully ready in onnxruntime).
    ('logreg', LogisticRegression()),
])

pipeline.fit(train_data, train.target)
print(classification_report(pipeline.predict(test_data), test.target))

              precision    recall  f1-score   support

           0       0.69      0.78      0.73       285
           1       0.75      0.66      0.70       285

    accuracy                           0.72       570
   macro avg       0.72      0.72      0.71       570
weighted avg       0.72      0.72      0.71       570

ONNX conversion

It is difficult to replicate the exact same tokenizer behaviour if the tokeniser comes from space, gensim or nltk. The default one used by scikit-learn uses regular expressions and is currently being implementing. The current implementation only considers a list of separators which can is defined in variable seps.

seps = {
    TfidfVectorizer: {
        "separators": [
            ' ', '.', '\\?', ',', ';', ':', '!',
            '\\(', '\\)', '\n', '"', "'",
            "-", "\\[", "\\]", "@"
        ]
    }
}
model_onnx = convert_sklearn(
    pipeline, "tfidf",
    initial_types=[("input", StringTensorType([None, 2]))],
    options=seps, target_opset=12)

And save.

with open("pipeline_tfidf.onnx", "wb") as f:
    f.write(model_onnx.SerializeToString())

Predictions with onnxruntime.

sess = rt.InferenceSession("pipeline_tfidf.onnx")
print('---', train_data[0])
inputs = {'input': train_data[:1]}
pred_onx = sess.run(None, inputs)
print("predict", pred_onx[0])
print("predict_proba", pred_onx[1])

--- [" Re: Jews can't hide from keith@cco."
 'Deletions...\n\nSo, you consider the german poster\'s remark anti-semitic?  Perhaps you\nimply that anyone in Germany who doesn\'t agree with israely policy in a\nnazi?  Pray tell, how does it even qualify as "casual anti-semitism"? \nIf the term doesn\'t apply, why then bring it up?\n\nYour own bigotry is shining through.  \n-- ']
predict [1]
predict_proba [{0: 0.4384377896785736, 1: 0.561562180519104}]

With scikit-learn:

print(pipeline.predict(train_data[:1]))
print(pipeline.predict_proba(train_data[:1]))

[0]
[[0.71903792 0.28096208]]

There are discrepencies for this model because the tokenization is not exactly the same. This is a work in progress.

Display the ONNX graph

Finally, let’s see the graph converted with sklearn-onnx.

pydot_graph = GetPydotGraph(
    model_onnx.graph, name=model_onnx.graph.name,
    rankdir="TB", node_producer=GetOpNodeProducer("docstring",
                                                  color="yellow",
                                                  fillcolor="yellow",
                                                  style="filled"))
pydot_graph.write_dot("pipeline_tfidf.dot")

os.system('dot -O -Gdpi=300 -Tpng pipeline_tfidf.dot')

image = plt.imread("pipeline_tfidf.dot.png")
fig, ax = plt.subplots(figsize=(40, 20))
ax.imshow(image)
ax.axis('off')

(-0.5, 4939.5, 11475.5, -0.5)