Vowpal-Wabbit - Scikit-Learn adaptor raises TypeError

[tdezhdar]

Describe the bug

Running the following python code

from vowpalwabbit.sklearn import VWClassifier
import numpy as np
import pandas as pd

X = pd.DataFrame({'a': [1, 2, 3, 4], 'b': [4, 1, 4, 2]}, dtype='uint32').astype('category')
y = pd.Series(np.zeros(4), dtype='int64')

VWClassifier(ftrl=True).fit(X, y)

gives a quite mysterious

Traceback (most recent call last):
  File "/xxx/scratch.py", line 8, in <module>
    VWClassifier(ftrl=True).fit(X, y)
  File "/xxx/python3.8/site-packages/vowpalwabbit/sklearn.py", line 568, in fit
    return VW.fit(self, X=X, y=y, sample_weight=sample_weight)
  File "/xxx/python3.8/site-packages/vowpalwabbit/sklearn.py", line 348, in fit
    X = tovw(
  File "/xxx/python3.8/site-packages/vowpalwabbit/sklearn.py", line 831, in tovw
    dump_svmlight_file(x, np.zeros(rows), s)
  File "/xxx/python3.8/site-packages/sklearn/datasets/_svmlight_format_io.py", line 525, in dump_svmlight_file
    _dump_svmlight(X, y, f, multilabel, one_based, comment, query_id)
  File "/xxx/python3.8/site-packages/sklearn/datasets/_svmlight_format_io.py", line 395, in _dump_svmlight
    _dump_svmlight_file(
  File "sklearn/datasets/_svmlight_format_fast.pyx", line 233, in sklearn.datasets._svmlight_format_fast._dump_svmlight_file
  File "sklearn/datasets/_svmlight_format_fast.pyx", line 131, in sklearn.datasets._svmlight_format_fast.__pyx_fused_cpdef
TypeError: Function call with ambiguous argument types

Note the odd original type of X -- this is output from another feature generator. It is possible to convert X first to int64 and then to category to circumvent the problem mostly. Only a cleanup error remains.

For the moment, it is a viable option to first convert X to a signed integer type and then to categorical, however that might not always be the case. Furthermore, this behaviour should probably be documented.

How to reproduce

See above.

Version

vowpalwabbit==9.8.0

OS

macos

Language

python

Additional context

No response

[jackgerrits]

Thanks for opening this. I looked into it and we're using a function in sklearn that seems to not work for unsigned types. I've added a more descriptive type error (#4610) and opened an issue (#4609) to track a real fix for this.

[tdezhdar]

Thanks @jackgerrits that more descriptive type error is already super useful.

[jackgerrits]

I'm going to go ahead and close this issue now that the better error is merged and the more complete fix will be tracked by #4609.