huggingface/pytorch-image-models · error · ValueError

Invalid local_mbconv_norm={local_mbconv_norm!r}; expected on

Error message

Invalid local_mbconv_norm={local_mbconv_norm!r}; expected one of {tuple(_LOCAL_MBCONV_NORM_MODES)}.

What it means

CPoTBone (cpubone.py) validates the local_mbconv_norm argument against a fixed set of modes (_LOCAL_MBCONV_NORM_MODES, e.g. 'none', 'first', 'last', 'all' style keys mapping to which MBConv sub-blocks get norm). Any string outside that tuple raises ValueError at construction time.

Source

Thrown at timm/models/cpubone.py:39

from ._features import feature_take_indices
from ._features_fx import register_notrace_module
from ._manipulate import checkpoint_seq
from ._registry import register_model, generate_default_cfgs

__all__ = ['CPUBone']


_LOCAL_MBCONV_NORM_MODES = {
    # mode: (expand, depthwise, project)
    'proj': (False, False, True),
    'depth_proj': (False, True, True),
    'all': (True, True, True),
}


def _check_local_mbconv_norm(local_mbconv_norm: str) -> None:
    if local_mbconv_norm not in _LOCAL_MBCONV_NORM_MODES:
        raise ValueError(
            f'Invalid local_mbconv_norm={local_mbconv_norm!r}; '
            f'expected one of {tuple(_LOCAL_MBCONV_NORM_MODES)}.'
        )


def _check_global_pool(global_pool: str) -> None:
    assert global_pool in ("", "avg"), "CPUBone only supports average or disabled pooling"


def remap_legacy_state_dict(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
    """Remap keys from original CPUBone checkpoints to the current model layout."""
    remapped = {}
    for k, v in state_dict.items():
        # conv_proj was nn.Sequential([conv, bn]) → now ConvLayer with .conv / .bn
        k = k.replace(".conv_proj.0.", ".conv_proj.conv.")
        k = k.replace(".conv_proj.1.", ".conv_proj.norm.")
        # pwise was a single-element nn.Sequential → now a plain nn.Conv2d
        k = k.replace(".pwise.0.", ".pwise.")

View on GitHub (pinned to 9a5261e31b)

Solutions

1. Print tuple(_LOCAL_MBCONV_NORM_MODES) from timm.models.cpubone and use one of those exact keys
2. Check the model's pretrained cfg / docs for the default value and revert to it
3. Update timm if the mode you want exists in a newer release

Example fix

# before
model = timm.create_model('cputbone_s', local_mbconv_norm='first-last')
# after
from timm.models.cpubone import _LOCAL_MBCONV_NORM_MODES
model = timm.create_model('cputbone_s', local_mbconv_norm=list(_LOCAL_MBCONV_NORM_MODES)[0])

from timm.models.cpubone import _LOCAL_MBCONV_NORM_MODES
if cfg['local_mbconv_norm'] not in _LOCAL_MBCONV_NORM_MODES:
    raise ValueError(f"local_mbconv_norm must be one of {tuple(_LOCAL_MBCONV_NORM_MODES)}")
model = timm.create_model('cputbone_s', **cfg)

def is_valid_local_mbconv_norm(v: str) -> bool:
    from timm.models.cpubone import _LOCAL_MBCONV_NORM_MODES
    return v in _LOCAL_MBCONV_NORM_MODES

When it happens

Trigger: Instantiating cputbone_s or CPoTBone with local_mbconv_norm set to an unrecognized string (e.g. 'first-last', 'norm', typo like 'fisrt').

Common situations: Hand-writing config YAML/JSON for backbone experiments; copying a norm-mode name from a different model family; version drift if valid mode names changed between timm releases.

Related errors

• Input image must have positive dimensions, got H={height}, W
• All scheduled batch sizes must be positive integers.
• num_batches must be a positive integer when specified.
• A progressive schedule requires at least two choices.
• schedule_epochs must be a positive integer for a progressive

AI-assisted analysis of huggingface/pytorch-image-models@9a5261e31b (2026-08-27). Data as JSON: /api/errors/acbf6e705799bff0. Report an issue: GitHub.