huggingface/pytorch-image-models · error · ValueError

output_fmt='NCHW' requires a raw image (B, C, H, W) input.

Error message

output_fmt='NCHW' requires a raw image (B, C, H, W) input.

What it means

Gemma4VitEncoder.forward_intermediates can reshape intermediate token outputs into spatial NCHW maps, but the grid can only be recovered from position_ids derived from a raw image input. If reshape=True (output_fmt='NCHW') and the input was pre-patchified or otherwise not a 4D raw image, the reshape is impossible and ValueError is raised.

Source

Thrown at timm/models/gemma4_vit.py:960

        def _cb(i: int, y: torch.Tensor) -> None:
            if i in take_indices:
                intermediates.append(y)

        max_block_index = None
        if stop_early and not torch.jit.is_scripting():
            max_block_index = max_index

        x = self._encode(
            x,
            position_ids,
            padding_positions,
            block_callback=_cb,
            max_block_index=max_block_index,
        )

        if reshape:
            if raw_input_ndim != 4:
                raise ValueError("output_fmt='NCHW' requires a raw image (B, C, H, W) input.")
            # Recover grid from internal (x, y) position_ids max.
            B = position_ids.shape[0]
            pW = int(position_ids[..., 0].max().item()) + 1
            pH = int(position_ids[..., 1].max().item()) + 1
            intermediates = [y.reshape(B, pH, pW, -1).permute(0, 3, 1, 2).contiguous() for y in intermediates]

        if output_dict:
            result_dict: Dict[str, Any] = {'image_intermediates': intermediates}
            if not intermediates_only:
                result_dict['image_features'] = x
            result_dict['patch_valid'] = ~padding_positions
            return result_dict

        if intermediates_only:
            return intermediates

        return x, intermediates

View on GitHub (pinned to 9a5261e31b)

Solutions

1. Pass a raw 4D image when you need output_fmt='NCHW'
2. Or use output_fmt='NLC' for pre-patchified input and reshape externally using your known grid (from patch_coord maxima)
3. Compute the grid yourself: pH/pW = position_ids.max()+1 per axis, then reshape tokens to (B, pH, pW, C) and permute

Example fix

# before
feats = enc.forward_intermediates(patches, patch_coord=c, output_fmt='NCHW')
# after
feats = enc.forward_intermediates(patches, patch_coord=c, output_fmt='NLC')
maps = [y.reshape(B, pH, pW, -1).permute(0, 3, 1, 2) for y in feats]

raw = x.ndim == 4 and patch_coord is None
fmt = 'NCHW' if raw else 'NLC'
feats = enc.forward_intermediates(x, patch_coord=patch_coord, output_fmt=fmt)

def can_use_nchw_output(x: torch.Tensor) -> bool:
    return x.ndim == 4

try:
    feats = enc.forward_intermediates(x, patch_coord=c, output_fmt='NCHW')
except ValueError as e:
    if 'NCHW' in str(e):
        feats = enc.forward_intermediates(x, patch_coord=c, output_fmt='NLC')
    else:
        raise

When it happens

Trigger: Calling enc.forward_intermediates(patches, patch_coord=..., output_fmt='NCHW') with a 3D/5D pre-patchified input instead of a raw (B,C,H,W) image.

Common situations: Reusing the same forward_intermediates call for both raw-image and cached-token paths with a fixed output_fmt; writing generic feature-extraction utilities that always request NCHW.

Related errors

• Expected input ndim in (3, 4, 5); got {x.ndim}.
• patch_coord is required for pre-patchified input.
• Cannot pool {N} tokens with k={k}: N must be divisible by k^
• Image size ({H}, {W}) must be divisible by (patch_size * poo
• Gemma4VitEncoder does not support classification use cases.

AI-assisted analysis of huggingface/pytorch-image-models@9a5261e31b (2026-08-27). Data as JSON: /api/errors/9275252cf0670fe7. Report an issue: GitHub.