pola-rs/polars · error · TypeError

cannot convert DataFrame to {target} (mixed type columns res

Error message

cannot convert DataFrame to {target} (mixed type columns result in `object` dtype)\n{df.schema!r}

What it means

After converting a DataFrame to a numpy array for PyTorch/Jax (frame_to_numpy), if the resulting array has dtype object the values cannot be loaded into a typed tensor, so TypeError is raised with the full frame schema printed for diagnosis. Object dtype appears when the frame mixes incompatible column types (e.g. String alongside numeric) or contains object columns.

Source

Thrown at py-polars/src/polars/ml/utilities.py:29

    *,
    writable: bool,
    target: str,
    order: IndexOrder = "fortran",
) -> np.ndarray[Any, Any]:
    """Convert a DataFrame to a NumPy array for use with Jax or PyTorch."""
    for nm, tp in df.schema.items():
        if tp == List:
            msg = f"cannot convert List column {nm!r} to {target} (use Array dtype instead)"
            raise TypeError(msg) from None

    if df.width == 1 and df.schema.dtypes()[0] == Array:
        arr = df[df.columns[0]].to_numpy(writable=writable)
    else:
        arr = df.to_numpy(writable=writable, order=order)

    if arr.dtype == object:
        msg = f"cannot convert DataFrame to {target} (mixed type columns result in `object` dtype)\n{df.schema!r}"
        raise TypeError(msg)
    return arr

View on GitHub (pinned to df599052da)

Solutions

Select only numeric feature columns before converting: df.select(cs.numeric()).to_torch()
Cast to a common numeric dtype: df.select(pl.col(c).cast(pl.Float32) for c in cols)
Encode string/categorical columns first (e.g. to codes or one-hot), or exclude them
Read the schema printed in the error to find the offending non-numeric column

Example fix

# before
df.to_torch()  # frame has 'name': String next to numeric columns

# after
import polars.selectors as cs
df.select(cs.numeric()).cast(pl.Float32).to_torch()

Defensive patterns

Strategy: validation

Validate before calling

import polars.selectors as cs
numeric = df.select(cs.numeric())
if numeric.width != df.width:
    bad = [c for c in df.columns if c not in numeric.columns]
    raise TypeError(f'non-numeric columns block tensor conversion: {bad}')

Type guard

import polars as pl
import polars.selectors as cs

def all_numeric(df: pl.DataFrame) -> bool:
    return df.width == df.select(cs.numeric()).width

Prevention

Select and cast features explicitly before to_torch/to_jax; never pass raw inferred frames
Encode labels (strings) to integers before tensor conversion

When it happens

Trigger: df.to_torch() / df.to_jax(return_type='array') where the frame has heterogeneous column dtypes (strings + numbers), pl.Object columns, or unconverted categorical/string features.

Common situations: Feeding raw CSV/inferred frames to ML conversion without selecting/casting features; string label columns left in the frame; mixed-type columns from dirty data.

Related errors

AI-assisted analysis of pola-rs/polars@df599052da (2026-08-16). Data as JSON: /api/errors/50311348372f0197. Report an issue: GitHub.