keras-team/keras · error · ValueError

Architecture configuration does not match {weights_name} var

Error message

Architecture configuration does not match {weights_name} variant. When using pre-trained weights, the model architecture must match the pre-trained configuration exactly. Expected depths: {expected_config['depths']}, got: {depths}. Expected projection_dims: {expected_config['projection_dims']}, got: {projection_dims}.

What it means

compute_output_spec of reconstruct_patches rejects a patches tensor whose rank matches neither the unbatched nor the batched layout for the chosen 2D/3D mode. Reconstruction must know whether the flat patch grid is (rows, cols[, planes], patch_flat) or has a leading batch dim; any other rank makes grid inference impossible.

Source

Thrown at keras/src/applications/convnext.py:533

        )(x)

    else:
        if pooling == "avg":
            x = layers.GlobalAveragePooling2D()(x)
        elif pooling == "max":
            x = layers.GlobalMaxPooling2D()(x)
        x = layers.LayerNormalization(epsilon=1e-6)(x)

    model = Functional(inputs=inputs, outputs=x, name=name)

    # 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":

View on GitHub (pinned to 7a34a03db6)

Solutions

Reshape patches to (grid_h, grid_w, patch_flat) or (batch, grid_h, grid_w, patch_flat) for 2D (add grid_d for 3D) before reconstruct_patches
Regenerate patches with extract_patches and pass them through unmodified
Check patches.ndim and the is_3d flag agree before the call

Example fix

before: recon = reconstruct_patches(tokens, size=(8,8)) where tokens.shape == (n, 192) -> ValueError; after: recon = reconstruct_patches(tokens.reshape(1, gh, gw, 192), size=(8,8))

Defensive patterns

Strategy: validation

Validate before calling

want = 5 if is_3d else 4
assert patches.ndim in (want - 1, want), f"patches rank {patches.ndim} not in {(want-1, want)}"

Type guard

def has_reconstructable_rank(patches, is_3d: bool) -> bool:
    batched, unbatched = (5, 4) if is_3d else (4, 3)
    return patches.ndim in (batched, unbatched)

Try / catch

try:
    recon = keras.ops.image.reconstruct_patches(patches, size=size)
except ValueError as e:
    raise ValueError(f"expected grid-shaped patches, got shape {patches.shape}") from e

Prevention

Keep the grid axes produced by extract_patches; only reshape the last axis
For ViT tokens, reshape (batch, tokens, dim) back to (batch, gh, gw, dim) first
Log patches.shape right before reconstruction when debugging pipelines

When it happens

Trigger: Passing a rank-3 tensor to a 2D reconstruction (grid dims lost, e.g. patches reshaped to (n_patches, flat) beforehand); passing a rank-5 tensor to 2D reconstruct; reconstructing 3D patches while the op was configured 2D (is_3d mismatch); extra/missing axes from squeeze/reshape in a preprocessing pipeline.

Common situations: Feeding patches straight from a dataloader that flattens them further; vision-transformer code that reshapes patch tokens to (batch, tokens, dim) and forgets to restore the 2D grid; switching a pipeline between 2D and 3D without regenerating the patches.

Related errors

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/bb5a0d6cbefd77b7. Report an issue: GitHub.