[BUG] validate argument does not work in imblearn.FunctionSampler

[raphaelberly]

Describe the bug

validate argument in imblearn.FunctionSampler does not seem to work. When the dataset contains null values, fit crashes even though validate is False and the sampler does not depend on dataset values.

Steps/Code to Reproduce

import pandas as pd
from imblearn import FunctionSampler

X = pd.DataFrame([{'a': 1, 'b': 1}, {'a': 1, 'b': None}])
y = pd.Series([1, 0])


def func(X, y):
    return X[:1], y[:1]


sampler = FunctionSampler(func=func, validate=False)
sampler.fit(X, y)

Expected Results

No error is thrown.

Actual Results

Traceback (most recent call last):
  File "...", line 13, in <module>
    sampler.fit(X, y)
  File ".../site-packages/imblearn/base.py", line 48, in fit
    X, y, _ = self._check_X_y(X, y)
  File ".../site-packages/imblearn/base.py", line 135, in _check_X_y
    X, y, reset=True, accept_sparse=accept_sparse
  File ".../site-packages/sklearn/base.py", line 432, in _validate_data
    X, y = check_X_y(X, y, **check_params)
  File ".../site-packages/sklearn/utils/validation.py", line 72, in inner_f
    return f(**kwargs)
  File ".../site-packages/sklearn/utils/validation.py", line 802, in check_X_y
    estimator=estimator)
  File ".../site-packages/sklearn/utils/validation.py", line 72, in inner_f
    return f(**kwargs)
  File ".../site-packages/sklearn/utils/validation.py", line 645, in check_array
    allow_nan=force_all_finite == 'allow-nan')
  File ".../site-packages/sklearn/utils/validation.py", line 99, in _assert_all_finite
    msg_dtype if msg_dtype is not None else X.dtype)
ValueError: Input contains NaN, infinity or a value too large for dtype('float64').

Versions

System:
    python: 3.7.7 (default, Mar 10 2020, 15:43:03)  [Clang 11.0.0 (clang-1100.0.33.17)]
   machine: Darwin-19.6.0-x86_64-i386-64bit

Python dependencies:
          pip: 20.2.3
   setuptools: 46.1.2
      sklearn: 0.23.2
        numpy: 1.18.4
        scipy: 1.4.1
       Cython: None
       pandas: 1.1.3
   matplotlib: 3.2.1
       joblib: 0.14.1
threadpoolctl: 2.0.0

Built with OpenMP: True

[glemaitre]

Yep this seems to be a regression. Thanks for reporting.

[glemaitre]

Oh actually I did not notice at first but actually you should only call fit_resample.

There is still a bug on our side because fit should be stateless with the FunctionSampler for sure. Only fit_resample should do something (which it does).

[raphaelberly]

Thanks for your response, @glemaitre. You are right, indeed, I should have called fit_resample.

The code I had that initially broke was using a pipeline:

import pandas as pd
from imblearn import FunctionSampler
from imblearn.pipeline import Pipeline

X = pd.DataFrame([{'a': 1, 'b': 1}, {'a': 1, 'b': None}])
y = pd.Series([1, 0])

def func(X, y):
    return X[:1], y[:1]

sampler = FunctionSampler(func=func, validate=False)
pipeline = Pipeline([('sampler', sampler)])

# The following line breaks
pipeline.fit(X, y)

If my understanding of your API is correct, that's how I should use the pipeline. This seems to confirm your conclusion, which is that there is still a bug on your side.

[glemaitre]

yes it happens here because the sampler is at the last position of the pipeline. They are fishy stuff here that need to be fix for sure.