OrdinalEncoder is not converting correctly when min_frequency is set

[chapman73]

As per the title. Code to reproduce is below:

import random

import numpy as np
import pandas as pd
from onnxruntime import InferenceSession
from skl2onnx import to_onnx
from skl2onnx.common.data_types import FloatTensorType
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import OrdinalEncoder

k = 100
sample_data = random.sample((list(range(1, 10)) + [np.nan]) * 1000, k=100)


samples = pd.DataFrame(
    {
        "sample_data": np.random.choice(range(100), size=k).tolist(),
        "target": np.random.choice([0, 1], size=k).tolist(),
    }
)

samples["sample_data"] = samples["sample_data"].astype("float32")


col = "sample_data"
X = samples[[col]]
y = samples["target"]
X_train, X_val, y_train, y_val = train_test_split(
    X, y, train_size=0.7, random_state=42, shuffle=True
)
X_val.dropna(subset=[col], inplace=True)

params = [{
    "handle_unknown": "use_encoded_value",
    "unknown_value": -1,
    "min_frequency": 5,
}, {
    "handle_unknown": "use_encoded_value",
    "unknown_value": -1,
}]


def test_model(params):
    model = OrdinalEncoder(**params)
    model.fit(X_train.values.astype(np.float32), y_train)


    model_onnx = to_onnx(
        model, initial_types=[("X", FloatTensorType([None, 1]))], target_opset=17
    )


    session = InferenceSession(model_onnx.SerializeToString())

    output = session.run(None, {"X": samples[[col]].values.astype(np.float32)})[0]
    print(output.flatten())
    # show which rows differ
    diff = np.abs(output - model.transform(samples[[col]].values))
    print(
        "With params", params,
        "There are",
        len(diff[diff > 0.0001]),
        "rows that differ from",
        len(samples[[col]].values),
    )


test_model(params[0])
test_model(params[1])

Sample output -- as you can see the output is the same even though min_frequency is set in one of the tests. This suggests that the parameter is not being respected.

$ python -i ordinal_encoder_only.py
[21. 17. 24. 11. -1. 21. 32. 47. 14. -1. 47. 44. 46. 47.  2. 33. 45. 20.
 -1.  4. 39. 30. 45. 23. 15. 43. -1.  3. 44. 21. 15. 45. 29. 35. 33. 42.
  7.  6. 35. 18. 29. 33.  6. 26.  4. -1. 20.  8.  1. 25.  0. 37. 39. 44.
 20. -1. 25. 12. 40. 35. 27. 19. 44. 21. 22.  5. 18. 43. 33. -1. 40. 31.
 -1. -1. 22. 29. -1.  3. 36. 38. 42. 34. 28. -1. 41. 46. 13. 28. -1. 31.
  0.  9. 10. 45. 16. 13.  9. 11. 26. 26.]
With params {'handle_unknown': 'use_encoded_value', 'unknown_value': -1, 'min_frequency': 5} There are 86 rows that differ from 100
[21. 17. 24. 11. -1. 21. 32. 47. 14. -1. 47. 44. 46. 47.  2. 33. 45. 20.
 -1.  4. 39. 30. 45. 23. 15. 43. -1.  3. 44. 21. 15. 45. 29. 35. 33. 42.
  7.  6. 35. 18. 29. 33.  6. 26.  4. -1. 20.  8.  1. 25.  0. 37. 39. 44.
 20. -1. 25. 12. 40. 35. 27. 19. 44. 21. 22.  5. 18. 43. 33. -1. 40. 31.
 -1. -1. 22. 29. -1.  3. 36. 38. 42. 34. 28. -1. 41. 46. 13. 28. -1. 31.
  0.  9. 10. 45. 16. 13.  9. 11. 26. 26.]
With params {'handle_unknown': 'use_encoded_value', 'unknown_value': -1} There are 0 rows that differ from 100

Any pointers to the best way to address this?

Thanks in advance.

[xadupre]

You can look at this PR #1195. I usually use the class ReferenceEvaluator from onnx and prints out the intermediate results in scikit-learn to understand when the divergence starts.