keras-team/keras · error · ValueError

Arguments `target` and `output` must have the same shape up…

Error message

Arguments `target` and `output` must have the same shape up until the last dimension: target.shape={target.shape}, output.shape={output.shape}

What it means

Raised by the NumPy backend's categorical crossentropy implementation when the shape guard that compares `target` and `output` finds they disagree on any dimension except the last one. The last axis of `output` holds class probabilities while `target` may be one-hot along that same axis, so all leading (batch/intermediate) dimensions must line up exactly; a mismatch means the labels tensor does not broadcast against the predictions tensor.

Solutions

1. Reshape or re-encode `target` so every dimension except the last matches `output` (e.g. one-hot encode integer labels with keras.ops.one_hot to the same class count)
2. Remove a spurious trailing axis of size 1 on `target` (squeeze it) when labels were stored as (..., 1) instead of (..., num_classes)
3. Check for accidental broadcasting upstream in the model that changed the batch or intermediate dimensions of the predictions before the loss was computed

When it happens

Trigger: Thrown at keras/src/backend/numpy/nn.py:899 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/437102ea6c7e35ea. Report an issue: GitHub.

Appendix: source

Thrown at keras/src/backend/numpy/nn.py:899

        output = np.clip(output, backend.epsilon(), 1.0 - backend.epsilon())
        log_prob = np.log(output)
    return -np.sum(target * log_prob, axis=axis)


def sparse_categorical_crossentropy(target, output, from_logits=False, axis=-1):
    target = np.array(target, dtype="int32")
    output = np.array(output)
    if len(target.shape) == len(output.shape) and target.shape[-1] == 1:
        target = np.squeeze(target, axis=-1)

    if len(output.shape) < 1:
        raise ValueError(
            "Argument `output` must be at least rank 1. "
            "Received: "
            f"output.shape={output.shape}"
        )
    if target.shape != output.shape[:-1]:
        raise ValueError(
            "Arguments `target` and `output` must have the same shape "
            "up until the last dimension: "
            f"target.shape={target.shape}, output.shape={output.shape}"
        )
    if from_logits:
        log_prob = log_softmax(output, axis=axis)
    else:
        output = output / np.sum(output, axis, keepdims=True)
        output = np.clip(output, backend.epsilon(), 1.0 - backend.epsilon())
        log_prob = np.log(output)
    target = one_hot(target, output.shape[axis], axis=axis)
    return -np.sum(target * log_prob, axis=axis)


def binary_crossentropy(target, output, from_logits=False):
    target = np.array(target)
    output = np.array(output)

View on GitHub (pinned to 7a34a03db6)