keras-team/keras · error · ValueError

`adapt()` can only be called on a tf.data.Dataset or a dict

Error message

`adapt()` can only be called on a tf.data.Dataset or a dict of arrays/lists. Received instead: {dataset} (of type {type(dataset)})

What it means

FeatureSpace.adapt() only accepts a tf.data.Dataset or a plain dict of arrays/lists so it can iterate samples and compute vocabulary statistics. Any other input type (DataFrame, list of dicts, numpy array, generator) is rejected immediately. Convert your data to one of the two accepted forms before adapting.

Source

Thrown at keras/src/layers/preprocessing/feature_space.py:516

            # Special case: a Normalization layer with preset mean/variance.
            # Not adaptable.
            if isinstance(preprocessor, layers.Normalization):
                if preprocessor.input_mean is not None:
                    continue
            # Special case: a TextVectorization layer with provided vocabulary.
            elif isinstance(preprocessor, layers.TextVectorization):
                if preprocessor._has_input_vocabulary:
                    continue
            if hasattr(preprocessor, "adapt"):
                adaptable_preprocessors.append(name)
        return adaptable_preprocessors

    def adapt(self, dataset, verbose=1):
        if not isinstance(dataset, tf.data.Dataset):
            if isinstance(dataset, dict):
                dataset = tf.data.Dataset.from_tensor_slices(dataset)
            else:
                raise ValueError(
                    "`adapt()` can only be called on a tf.data.Dataset or a "
                    "dict of arrays/lists. "
                    f"Received instead: {dataset} (of type {type(dataset)})"
                )

        adaptable_preprocessors = self._list_adaptable_preprocessors()
        if not adaptable_preprocessors:
            self._is_adapted = True
            self.get_encoded_features()
            self.built = True
            self._sublayers_built = True
            return

        # Check if the dataset needs batching
        x = next(iter(dataset))
        if len(x[adaptable_preprocessors[0]].shape) == 0:
            dataset = dataset.batch(32)

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Convert the input to a dict of arrays: fs.adapt({'col1': arr1, 'col2': arr2})
2. Or wrap it in a dataset: fs.adapt(tf.data.Dataset.from_tensor_slices(dict(df)))
3. For DataFrames: fs.adapt({c: df[c].values for c in df.columns})

Example fix

// before
fs.adapt(df)  # pandas DataFrame
// after
fs.adapt({col: df[col].values for col in df.columns})
# or: fs.adapt(tf.data.Dataset.from_tensor_slices(dict(df)))

if not isinstance(dataset, tf.data.Dataset) and not isinstance(dataset, dict):
    raise TypeError("adapt() needs a tf.data.Dataset or dict of arrays")

def is_adaptable(x):
    return isinstance(x, tf.data.Dataset) or (
        isinstance(x, dict) and all(hasattr(v, "__len__") or hasattr(v, "__getitem__") for v in x.values())
    )

catch ValueError around fs.adapt(raw_data) and fall back to converting dict inputs to a tf.data.Dataset before retrying

When it happens

Trigger: fs.adapt(pandas_dataframe), fs.adapt(numpy_array), fs.adapt(list_of_dicts), or fs.adapt(generator) — anything that is not a tf.data.Dataset or dict.

Common situations: Passing a pandas DataFrame, a numpy array, or a list of dicts to adapt(). Loading data with pd.read_csv and handing it straight to FeatureSpace.

Related errors

• A FeatureSpace can only be called with a dict. Received: dat
• Feature '{name}' has `output_mode='one_hot'`. Thus its prepr
• Feature '{name}' has `output_mode='one_hot'`. However it isn
• Cannot concatenate features because feature '{name}' has not
• You need to call `.adapt(dataset)` on the FeatureSpace befor

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/ea97a3d0ee3c17e6. Report an issue: GitHub.