keras-team/keras · error · ValueError

Model name "{name}" does not match weights variant "{weights

Error message

Model name "{name}" does not match weights variant "{weights_name}". When using imagenet weights, model name must contain the weights variant (e.g., "convnext_{weights_name.split("convnext_")[-1]}").

What it means

compute_output_spec of reconstruct_patches requires the last axis of patches to be exactly prod(size)*channels (the flattened patch content). If the final dim is not divisible by patch_d*patch_h*patch_w, the flattened values cannot be reshaped back into spatial patches, so it fails before any computation.

Source

Thrown at keras/src/applications/convnext.py:543

    # Validate weights before requesting them from the API
    if weights == "imagenet":
        expected_config = MODEL_CONFIGS[weights_name.split("convnext_")[-1]]
        if (
            depths != expected_config["depths"]
            or projection_dims != expected_config["projection_dims"]
        ):
            raise ValueError(
                f"Architecture configuration does not match {weights_name} "
                f"variant. When using pre-trained weights, the model "
                f"architecture must match the pre-trained configuration "
                f"exactly. Expected depths: {expected_config['depths']}, "
                f"got: {depths}. Expected projection_dims: "
                f"{expected_config['projection_dims']}, got: {projection_dims}."
            )

        if weights_name not in name:
            raise ValueError(
                f'Model name "{name}" does not match weights variant '
                f'"{weights_name}". When using imagenet weights, model name '
                f'must contain the weights variant (e.g., "convnext_'
                f'{weights_name.split("convnext_")[-1]}").'
            )

    # Load weights.
    if weights == "imagenet":
        if include_top:
            file_suffix = ".h5"
            file_hash = WEIGHTS_HASHES[weights_name][0]
        else:
            file_suffix = "_notop.h5"
            file_hash = WEIGHTS_HASHES[weights_name][1]
        file_name = name + file_suffix
        weights_path = file_utils.get_file(
            file_name,
            BASE_WEIGHTS_PATH + file_name,

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Ensure patches.shape[-1] == prod(size) * channels; if a linear layer changed the dim, project back to prod(size)*channels first
2. Pass exactly the same size (and data_format) used for extract_patches
3. If reconstructing a projected embedding, invert the projection or skip reconstruction

Example fix

before: reconstruct_patches(x, size=(8,8)) with x.shape[-1] == 768 and 3 channels; after: x2 = ops.matmul(x, W_inv); reconstruct_patches(x2, size=(8,8)) where x2.shape[-1] == 192

import math
prod_size = size if isinstance(size, int) else math.prod(size)
assert patches.shape[-1] % prod_size == 0, (patches.shape[-1], prod_size)

def flat_dim_matches(flat: int, size, channels: int) -> bool:
    prod_size = size if isinstance(size, int) else math.prod(size)
    return flat % (prod_size * channels) == 0 and flat > 0

When it happens

Trigger: Passing a projection/output dim of a ViT layer (e.g. 768) where the patch flatten dim is 8*8*3=192; using size=(8,8) when patches were extracted with (4,4); channels_first patches fed without specifying data_format, so C ends up on the wrong axis.

Common situations: Vision-transformer pipelines that append a dense projection after patch extraction and reconstruct from the projected tensor; patch size changed between extract and reconstruct; forgetting that the flat dim includes channels for channels_last data.

Related errors

• Architecture configuration does not match {weights_name} var
• DenseNet does not support the `channels_first` image data fo
• Unknown activation function '{activation}' cannot be seriali
• Could not interpret activation function identifier: {identif
• ConvNeXt does not support the `channels_first` image data fo

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/123a0aa4bd842130. Report an issue: GitHub.