keras-team/keras · error · ValueError

Feature '{name}' has `output_mode='one_hot'`. Thus its prepr

Error message

Feature '{name}' has `output_mode='one_hot'`. Thus its preprocessor should return an integer dtype. Instead it returns a {dtype} dtype.

What it means

When FeatureSpace is configured to produce one-hot output for a feature (output_mode='one_hot'), the feature's preprocessor must emit integer indices that can be binarized. The code checks the dtype flowing out of the preprocessor and rejects non-integer output such as float32 or string.

Source

Thrown at keras/src/layers/preprocessing/feature_space.py:685

        ] + [
            self.crosses_by_name[name] for name in self._crossed_features_names
        ]

        for name, feature, spec in zip(all_names, all_features, all_specs):
            if tree.is_nested(feature):
                dtype = tree.flatten(feature)[0].dtype
            else:
                dtype = feature.dtype
            dtype = backend.standardize_dtype(dtype)

            if spec.output_mode == "one_hot":
                preprocessor = self.preprocessors.get(
                    name
                ) or self.crossers.get(name)

                cardinality = None
                if not dtype.startswith("int"):
                    raise ValueError(
                        f"Feature '{name}' has `output_mode='one_hot'`. "
                        "Thus its preprocessor should return an integer dtype. "
                        f"Instead it returns a {dtype} dtype."
                    )

                if isinstance(
                    preprocessor, (layers.IntegerLookup, layers.StringLookup)
                ):
                    cardinality = preprocessor.vocabulary_size()
                elif isinstance(preprocessor, layers.CategoryEncoding):
                    cardinality = preprocessor.num_tokens
                elif isinstance(preprocessor, layers.Discretization):
                    cardinality = preprocessor.num_bins
                elif isinstance(
                    preprocessor, (layers.HashedCrossing, layers.Hashing)
                ):
                    cardinality = preprocessor.num_bins
                else:

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Make the feature's preprocessor return integer indices (e.g. use integer_hashed, string_hashed, or a lookup layer with output_mode='int')
2. If the feature must stay continuous, do not one-hot encode it — use output_mode='float' for that feature or keep it out of one_hot mode
3. For crossing features in one_hot mode, set crossing_output_mode='one_hot' or 'int' with an appropriate hasher

Example fix

// before
fs = FeatureSpace(..., output_mode="one_hot")
# feature 'x' preprocessor returns float
// after
fs = FeatureSpace(
    features={"x": FeatureSpace.integer_hashed(max_tokens=32)},
    output_mode="one_hot",
)

out = fs.preprocessors[name](sample_batch)
assert str(out.dtype).startswith("int"), f"{name} yields {out.dtype}"

def is_int_preprocessor(p):
    d = getattr(p, "dtype", None)
    return d is not None and str(d).startswith("int")

catch ValueError from fs.get_encoded_features() and inspect the offending feature's preprocessor output dtype via fs.preprocessors[name] before correcting the feature spec

When it happens

Trigger: A feature whose preprocessor returns floats (e.g. normalization) combined with FeatureSpace output_mode='one_hot'; a crossing outputting floats while crossing_output_mode='one_hot'.

Common situations: A custom feature spec or crossing returns floats; mixing output_mode='one_hot' with a preprocessor that has not been set to integer output; misconfigured crossing_output_mode on FeatureSpace.

Related errors

• Feature '{name}' has `output_mode='one_hot'`. However it isn
• `dtype` was passed both positionally and as a keyword argume
• Quantization mode='{mode}' doesn't work well with compute_dt
• `adapt()` can only be called on a tf.data.Dataset or a dict
• Cannot concatenate features because feature '{name}' has not

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/59b87ea87de2c6ed. Report an issue: GitHub.