keras-team/keras · error · ValueError

`patches` last dim ({static_flat}) is not divisible by prod(

Error message

`patches` last dim ({static_flat}) is not divisible by prod(size) ({pD * pH * pW}). Are `size` and the patches tensor consistent?

What it means

In 3D reconstruction the channels count C is inferred by dividing the patches' last (flat) dimension by prod(size)=pD*pH*pW. If flat is not divisible by the patch volume, size and the patches tensor disagree, so C would not be an integer and it raises.

Source

Thrown at keras/src/ops/image.py:1380

        raise ValueError(
            "`patches` has unexpected rank for 3D reconstruction. "
            "Expected 4 (unbatched) or 5 (batched). "
            f"Received shape: {patches.shape}"
        )

    _unbatched = False
    if len(patches.shape) == 4:
        _unbatched = True
        patches = backend.numpy.expand_dims(patches, axis=0)

    shp = ops.shape(patches)
    B, gD, gH, gW = shp[0], shp[1], shp[2], shp[3]
    static_flat = patches.shape[-1]
    if static_flat is None:
        C = shp[4] // (pD * pH * pW)
    else:
        if static_flat % (pD * pH * pW) != 0:
            raise ValueError(
                f"`patches` last dim ({static_flat}) is not divisible by "
                f"prod(size) ({pD * pH * pW}). Are `size` and the patches "
                f"tensor consistent?"
            )
        C = static_flat // (pD * pH * pW)

    x = backend.numpy.reshape(patches, (B, gD, gH, gW, pD, pH, pW, C))
    x = backend.numpy.transpose(x, axes=(0, 1, 4, 2, 5, 3, 6, 7))
    x = backend.numpy.reshape(x, (B, gD * pD, gH * pH, gW * pW, C))

    if padding == "same":
        static_gD = patches.shape[1]
        static_gH = patches.shape[2]
        static_gW = patches.shape[3]
        if isinstance(static_gD, int) and not (
            static_gD * pD - pD < D <= static_gD * pD
        ):
            raise ValueError(

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Pass exactly the same size used in extract_patches
2. Verify patches.shape[-1] == C * pD * pH * pW; solve for size if C is known
3. If extra dims were concatenated into the flat dim, split them out first

Example fix

# before
reconstruct_patches(patches, size=(2,8,8), output_size=out)
# flat=768, 2*8*8=128 -> 768%128==0 ok; but with size=(4,8,8): 768%256!=0

# after
size = (4, 8, 8)  # match extraction; flat=1024 for C=4
reconstruct_patches(patches, size=size, output_size=out)

flat = patches.shape[-1]
assert flat is None or flat % (size[0]*size[1]*size[2]) == 0, 'size/patches mismatch'

def size_matches_patches(patches, size) -> bool:
    flat = patches.shape[-1]
    return flat is None or flat % (size[0]*size[1]*size[2]) == 0

When it happens

Trigger: reconstruct_patches(patches, size=(pD,pH,pW), ...) where patches.shape[-1] % (pD*pH*pW) != 0, e.g. size=(4,8,8) (256) but flat dim 192, or passing a 2D-consistent size for 3D patches.

Common situations: Using a different patch size at reconstruction than at extraction; flattening extra features into the last dim; patches produced by a different layer/version with a different patching scheme.

Related errors

• For `padding='same'`, `output_size` width ({W}) must be in t
• `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 images rank: expected rank 3 (single image) or rank

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/ae5152b949189373. Report an issue: GitHub.