keras-team/keras · error · ValueError

If using `weights="imagenet"` with `include_top=True`, `clas

Error message

If using `weights="imagenet"` with `include_top=True`, `classes` should be 1000. Received classes={classes}

What it means

Raised by _extract_patches_3d when strides, after int expansion, is not a length-3 sequence. Strides default to the patch size; if you override them you must give one stride per spatial dim (d, h, w) so the extractor knows the sampling step along each axis.

Source

Thrown at keras/src/applications/convnext.py:409

    Returns:
        A model instance.
    """
    if backend.image_data_format() == "channels_first":
        raise ValueError(
            "ConvNeXt does not support the `channels_first` image data "
            "format. Switch to `channels_last` by editing your local "
            "config file at ~/.keras/keras.json"
        )
    if not (weights in {"imagenet", None} or file_utils.exists(weights)):
        raise ValueError(
            "The `weights` argument should be either "
            "`None` (random initialization), `imagenet` "
            "(pre-training on ImageNet), "
            "or the path to the weights file to be loaded."
        )

    if weights == "imagenet" and include_top and classes != 1000:
        raise ValueError(
            'If using `weights="imagenet"` with `include_top=True`, '
            "`classes` should be 1000. "
            f"Received classes={classes}"
        )

    # Determine proper input shape.
    input_shape = imagenet_utils.obtain_input_shape(
        input_shape,
        default_size=default_size,
        min_size=32,
        data_format=backend.image_data_format(),
        require_flatten=include_top,
        weights=weights,
    )

    if input_tensor is None:
        img_input = layers.Input(shape=input_shape)
    else:

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Pass strides as an int (same stride on all 3 axes) or a length-3 tuple matching (depth, height, width)
2. Omit strides entirely if you want non-overlapping patches (defaults to size)
3. Mirror the shape of size when building strides programmatically: strides = tuple(s // 2 for s in size)

Example fix

before: extract_patches_3d(v, size=(4,4,4), strides=(2,2)) -> ValueError; after: extract_patches_3d(v, size=(4,4,4), strides=2)

if strides is not None and not isinstance(strides, int):
    strides = tuple(strides)
    assert len(strides) == 3, f"strides for 3D must have length 3, got {len(strides)}"

def valid_3d_strides(strides) -> bool:
    return strides is None or isinstance(strides, int) or (isinstance(strides, (tuple, list)) and len(strides) == 3)

When it happens

Trigger: extract_patches_3d(vols, size=(4,4,4), strides=(2,2)) or strides=[2] on 5D inputs; passing a 2D stride tuple from an image pipeline into a volume pipeline; strides computed as (s, s) + something that dropped an axis.

Common situations: Reusing an image-model stride config for video/voxel models; strides derived from size[:-1] or another slicing bug; frameworks that accept 2-tuples elsewhere (conv2d) making the 3-tuple requirement easy to miss.

Related errors

• Unknown activation function '{activation}' cannot be seriali
• Could not interpret activation function identifier: {identif
• ConvNeXt does not support the `channels_first` image data fo
• The `weights` argument should be either `None` (random initi
• If using `weights` as `"imagenet"` with `include_top` as tru

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/f1d3c36e2a534769. Report an issue: GitHub.