Fix dtype of empty feature shuffle patterns

[Echo-Nie]

Problem

Thx for open-sourcing TabFM — really appreciate the work!

While running TabFM on the KC1 dataset with an n-shot retrieved context, predict_proba crashes with:

IndexError: arrays used as indices must be of integer (or boolean) type

The relevant traceback is:

File "tabfm/src/classifier_and_regressor.py", line 1630, in _transform_features
    shuffled_cols = X_variant_instance[:, shuffle_pattern]
IndexError: arrays used as indices must be of integer (or boolean) type

The shuffle patterns are converted with:

return [np.array(p) for p in shuffle_patterns]

When p is empty, NumPy infers float64. The resulting array is later used for column indexing, which requires an integer or boolean dtype.

Fix

Convert shuffle patterns using NumPy's native indexing dtype:

- return [np.array(p) for p in shuffle_patterns]
+ return [np.asarray(p, dtype=np.intp) for p in shuffle_patterns]

This preserves the values and ordering of non-empty patterns while making empty patterns valid NumPy index arrays.

Reproduce

import numpy as np

shuffle_pattern = np.array([])
X = np.empty((3, 0))

print(shuffle_pattern.dtype)  # float64
X[:, shuffle_pattern]         # IndexError before this change

After this change:

shuffle_pattern = np.asarray([], dtype=np.intp)
X = np.empty((3, 0))

result = X[:, shuffle_pattern]
print(result.shape)  # (3, 0)

Test

Ran the existing classifier and regressor tests:

pytest -q tabfm/src/classifier_and_regressor_test.py

18 passed, 28 skipped in 3.06s

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[Echo-Nie]

The PR is ready to be reviewed, thx!
@weihaokong