keras-team/keras · error · ValueError

compute_iou() expects boxes2 to be batched, or to be unbatch

Error message

compute_iou() expects boxes2 to be batched, or to be unbatched. Received len(boxes1.shape)={boxes1_rank}, len(boxes2.shape)={boxes2_rank}. Expected either len(boxes2.shape)=2 AND or len(boxes2.shape)=3.

What it means

Same rank contract as boxes1, applied to the second argument: compute_iou requires boxes2 to be rank 2 ((M, 4), unbatched) or rank 3 ((batch, M, 4), batched). A boxes2 of any other rank raises this ValueError with both ranks echoed in the message.

Source

Thrown at keras/src/layers/preprocessing/image_preprocessing/bounding_boxes/iou.py:112

            `image_shape` is used for normalization.

    Returns:
        iou_lookup_table: a vector containing the pairwise ious of boxes1 and
            boxes2.
    """  # noqa: E501

    boxes1_rank = len(ops.shape(boxes1))
    boxes2_rank = len(ops.shape(boxes2))

    if boxes1_rank not in [2, 3]:
        raise ValueError(
            "compute_iou() expects boxes1 to be batched, or to be unbatched. "
            f"Received len(boxes1.shape)={boxes1_rank}, "
            f"len(boxes2.shape)={boxes2_rank}. Expected either "
            "len(boxes1.shape)=2 AND or len(boxes1.shape)=3."
        )
    if boxes2_rank not in [2, 3]:
        raise ValueError(
            "compute_iou() expects boxes2 to be batched, or to be unbatched. "
            f"Received len(boxes1.shape)={boxes1_rank}, "
            f"len(boxes2.shape)={boxes2_rank}. Expected either "
            "len(boxes2.shape)=2 AND or len(boxes2.shape)=3."
        )

    target_format = "yxyx"
    if "rel" in bounding_box_format and image_shape is None:
        raise ValueError(
            "When using relative bounding box formats (e.g. `rel_yxyx`) "
            "the `image_shape` argument must be provided."
            f"Received `image_shape`: {image_shape}"
        )

    if image_shape is None:
        height, width = None, None
    else:
        height, width, _ = image_shape

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Ensure boxes2 has shape (M, 4) or (batch, M, 4).
2. Add a leading batch axis with ops.expand_dims(boxes2, axis=0) when boxes1 is batched.
3. Validate len(ops.shape(boxes2)) in [2, 3] before calling in data-pipeline code.

Example fix

# before
iou = compute_iou(pred_boxes, gt_flat_list)
# after
iou = compute_iou(pred_boxes, ops.convert_to_tensor(gt_flat_list).reshape(-1, 4))

boxes2 = ops.convert_to_tensor(boxes2)
assert len(ops.shape(boxes2)) in (2, 3), f"boxes2 rank {len(ops.shape(boxes2))} not in (2,3)"

def boxes_have_valid_rank(t):
    return len(ops.shape(ops.convert_to_tensor(t))) in (2, 3)

When it happens

Trigger: Passing ground-truth boxes as a flat list of coordinates, a rank-4 array, or a nesting depth that does not match a batched boxes1.

Common situations: Comparing predictions (batched) against labels loaded from a JSON that are flat or singly nested; forgetting to expand dims on the reference boxes.

Related errors

• compute_iou() expects boxes1 to be batched, or to be unbatch
• When using relative bounding box formats (e.g. `rel_yxyx`) t
• Architecture configuration does not match {weights_name} var
• `height` and `width` must be set if `format='xyxy'`.
• `variance` must be length 4, got {variance}

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/f0b83b17dc4ab51d. Report an issue: GitHub.