keras-team/keras · error · ValueError

All `HashedCrossing` inputs should be dense tensors. Receive

Error message

All `HashedCrossing` inputs should be dense tensors. Received: inputs={inputs}

What it means

HashedCrossing accepts only dense tensors. RaggedTensor or SparseTensor inputs are rejected because the crossing op needs uniform dense layout.

Source

Thrown at keras/src/layers/preprocessing/hashed_crossing.py:218

    def _check_input_shape_and_type(self, inputs):
        first_shape = tuple(inputs[0].shape)
        rank = len(first_shape)
        if rank > 2 or (rank == 2 and first_shape[-1] != 1):
            raise ValueError(
                "All `HashedCrossing` inputs should have shape `()`, "
                "`(batch_size)` or `(batch_size, 1)`. "
                f"Received: inputs={inputs}"
            )
        if not all(tuple(x.shape) == first_shape for x in inputs[1:]):
            raise ValueError(
                "All `HashedCrossing` inputs should have equal shape. "
                f"Received: inputs={inputs}"
            )
        if any(
            isinstance(x, (tf.RaggedTensor, tf.SparseTensor)) for x in inputs
        ):
            raise ValueError(
                "All `HashedCrossing` inputs should be dense tensors. "
                f"Received: inputs={inputs}"
            )
        if not all(
            tf.as_dtype(x.dtype).is_integer or x.dtype == tf.string
            for x in inputs
        ):
            raise ValueError(
                "All `HashedCrossing` inputs should have an integer or "
                f"string dtype. Received: inputs={inputs}"
            )

View on GitHub (pinned to 7a34a03db6)

Solutions

Convert ragged inputs with x.to_tensor() (possibly after trimming/padding to (batch,1))
Convert sparse inputs with tf.sparse.to_dense(x)
Move dense conversion upstream into the tf.data pipeline

Example fix

// before
out = layer([ragged_a, b])
// after
dense_a = ragged_a.to_tensor()
out = layer([dense_a, b])

Defensive patterns

Strategy: validation

Validate before calling

inputs = [x.to_tensor() if hasattr(x, "to_tensor") else tf.sparse.to_dense(x) if isinstance(x, tf.SparseTensor) else x for x in inputs]

Type guard

def all_dense(inputs):
    import tensorflow as tf
    return not any(isinstance(x, (tf.RaggedTensor, tf.SparseTensor)) for x in inputs)

Try / catch

catch ValueError from call() and densify inputs (x.to_tensor() / tf.sparse.to_dense(x)) before retrying

Prevention

Convert RaggedTensor/SparseTensor inputs to dense with to_tensor()/tf.sparse.to_dense before HashedCrossing
Do not pipe padded or ragged-batch datasets directly into a crossing layer

When it happens

Trigger: Any input x for which isinstance(x, (tf.RaggedTensor, tf.SparseTensor)) is true when the layer is called.

Common situations: Using ragged or sparse data pipelines (NLP token batches, padded sequences) directly with HashedCrossing without densifying.

Related errors

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/a235ad3d7fd3aea7. Report an issue: GitHub.