keras-team/keras · error · ValueError

For `padding='same'`, `output_size` width ({W}) must be in t

Error message

For `padding='same'`, `output_size` width ({W}) must be in the range ((gW-1)*pW, gW*pW], i.e. ({static_gW * pW - pW}, {static_gW * pW}]. Got: gW={static_gW}, pW={pW}.

What it means

In keras.ops.image reconstruct_patches (2D), padding='same' re-crops a padded reconstruction down to output_size. That crop is only consistent when the requested width W lies in ((gW-1)*pW, gW*pW], where gW is the patch-grid width and pW the patch width. Outside that window no 'same' extraction could have produced this grid for that image size, so the op raises.

Source

Thrown at keras/src/ops/image.py:1295

    x = backend.numpy.transpose(x, axes=(0, 1, 3, 2, 4, 5))
    x = backend.numpy.reshape(x, (B, gH * pH, gW * pW, C))

    if padding == "same":
        static_gH = patches.shape[1]
        static_gW = patches.shape[2]
        if isinstance(static_gH, int) and not (
            static_gH * pH - pH < H <= static_gH * pH
        ):
            raise ValueError(
                f"For `padding='same'`, `output_size` height ({H}) must "
                f"be in the range ((gH-1)*pH, gH*pH], i.e. "
                f"({static_gH * pH - pH}, {static_gH * pH}]. "
                f"Got: gH={static_gH}, pH={pH}."
            )
        if isinstance(static_gW, int) and not (
            static_gW * pW - pW < W <= static_gW * pW
        ):
            raise ValueError(
                f"For `padding='same'`, `output_size` width ({W}) must "
                f"be in the range ((gW-1)*pW, gW*pW], i.e. "
                f"({static_gW * pW - pW}, {static_gW * pW}]. "
                f"Got: gW={static_gW}, pW={pW}."
            )
        pad_total_h = gH * pH - H
        pad_total_w = gW * pW - W
        begin = [0, pad_total_h // 2, pad_total_w // 2, 0]
        out_shape = [B, H, W, C]
        x = ops.slice(x, begin, out_shape)
    else:
        if gH * pH != H or gW * pW != W:
            raise ValueError(
                f"`padding='valid'` requires output_size to equal "
                f"size * grid. Got output_size=({H},{W}), "
                f"grid=({gH},{gW}), size=({pH},{pW})."
            )

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Pass the exact width of the image that was originally patched (extract_patches input width) — it always satisfies the range
2. Otherwise pick any W in ((gW-1)*pW, gW*pW], e.g. the full grid width gW*pW
3. Recompute the grid: for same-padding extraction gW = ceil(W_original / pW); use that to sanity-check
4. If you need an output not covered by the grid, switch to padding='valid' with exact size*grid

Example fix

# before
reconstruct_patches(patches, size=(8,8), output_size=(20,20), padding='same')
# gW=4, pW=8 -> W must be in (24,32]

# after
reconstruct_patches(patches, size=(8,8), output_size=(20,28), padding='same')

gW, pW = int(patches.shape[2]), size[1]
assert (gW-1)*pW < output_size[1] <= gW*pW, 'W out of same-range'  # if gW static

def same_w_ok(patches, size, output_size) -> bool:
    gW = patches.shape[2]
    return not isinstance(gW, int) or ((gW-1)*size[1] < output_size[1] <= gW*size[1])

When it happens

Trigger: keras.ops.image.reconstruct_patches(patches, size, output_size, padding='same') with output_size width <= (gW-1)*pW or > gW*pW, e.g. gW=4, pW=8 but W=20 (needs 24<W<=32) or W=40.

Common situations: Reconstructing patches extracted with different padding/strides than assumed; passing a resized image size instead of the original extract-time width; swapping height/width in output_size; hardcoding output_size for a variable patch grid.

Related errors

• `padding='valid'` requires output_size to equal size * grid.
• For `padding='same'`, `output_size` depth ({D}) must be in t
• `padding='valid'` requires output_size to equal size * grid.
• Invalid `output_size`. Expected length 3 (D, H, W). Got: out
• `patches` has unexpected rank for 3D reconstruction. Expecte

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/5eab214b16e0ef50. Report an issue: GitHub.