keras-team/keras · error · ValueError

start_points and end_points must have the same shape. Receiv

Error message

start_points and end_points must have the same shape. Received start_points.shape={start_points.shape}, end_points.shape={end_points.shape}

What it means

perspective requires start_points and end_points to have identical shapes so each source corner maps to one destination corner. Mismatched batch sizes, or one array batched while the other is single-image, triggers this.

Source

Thrown at keras/src/ops/image.py:2194

    def compute_output_spec(self, images, start_points, end_points):
        if len(images.shape) not in (3, 4):
            raise ValueError(
                "Invalid images rank: expected rank 3 (single image) "
                "or rank 4 (batch of images). Received input with shape: "
                f"images.shape={images.shape}"
            )
        if start_points.shape[-2:] != (4, 2) or start_points.ndim not in (2, 3):
            raise ValueError(
                "Invalid start_points shape: expected (4,2) for a single image"
                f" or (N,4,2) for a batch. Received shape: {start_points.shape}"
            )
        if end_points.shape[-2:] != (4, 2) or end_points.ndim not in (2, 3):
            raise ValueError(
                "Invalid end_points shape: expected (4,2) for a single image"
                f" or (N,4,2) for a batch. Received shape: {end_points.shape}"
            )
        if start_points.shape != end_points.shape:
            raise ValueError(
                "start_points and end_points must have the same shape."
                f" Received start_points.shape={start_points.shape}, "
                f"end_points.shape={end_points.shape}"
            )
        return KerasTensor(images.shape, dtype=images.dtype)


@keras_export("keras.ops.image.perspective_transform")
def perspective_transform(
    images,
    start_points,
    end_points,
    interpolation="bilinear",
    fill_value=0,
    data_format=None,
):
    """Applies a perspective transformation to the image(s).

View on GitHub (pinned to 7a34a03db6)

Solutions

Make both arrays the same shape: either both (4, 2) or both (N, 4, 2) with equal N.
Tile/repeat the single-image array to batch size when it is constant.
Assert start_points.shape == end_points.shape before the call.

Example fix

# before
perspective(imgs, starts, ends[:4])  # ends has 8 entries

# after
ends = np.stack([base] * len(imgs))
perspective(imgs, starts, ends)

Defensive patterns

Strategy: validation

Validate before calling

import numpy as np
sp, ep = np.asarray(start_points), np.asarray(end_points)
if sp.ndim == 2 and ep.ndim == 3: sp = np.stack([sp] * ep.shape[0])
assert sp.shape == ep.shape, (sp.shape, ep.shape)

Type guard

def matching_corner_points(sp, ep):
    sp, ep = np.asarray(sp), np.asarray(ep)
    return sp.shape == ep.shape and sp.shape[-2:] == (4, 2)

Prevention

Broadcast constant corners explicitly with np.stack.
Keep source and destination corner lists generated in the same loop.

When it happens

Trigger: start_points.shape=(4,2) with end_points.shape=(8,4,2); a batch of N source corners but M != N destination corners.

Common situations: Broadcasting expectations carried over from NumPy conventions; per-image corners generated in a loop with an off-by-one; some images dropped from one list.

Related errors

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/9a903ab6cbe499eb. Report an issue: GitHub.