deepfakes/faceswap · error · ValueError

'{method}' is not a valid clipping method. Select from {list

Error message

'{method}' is not a valid clipping method. Select from {list(methods)}

What it means

Raised by the trainer's gradient-clipping factory in lib/training/optimizer.py when the clipping method string in the optimizer section of the training configuration does not match one of the implemented clippers: 'autoclip', 'global_norm', 'norm', or 'value'. It is a plain ValueError and surfaces before training starts, as soon as the optimizer is built.

Source

Thrown at lib/training/optimizer.py:125

        Parameters
        ----------
        method
            The clipping method to use
        autoclip_history
            The history length for auto clipping

        Returns
        -------
        The function used to clip the gradients
        """
        methods: dict[str, T.Callable[[list[nn.Parameter], float], None | torch.Tensor]] = {
            "autoclip": AutoClipper(int(self._value * 10), history_size=autoclip_history),
            "global_norm": nn.utils.clip_grad_norm_,
            "norm": self._clip_norm,
            "value": nn.utils.clip_grad_value_}
        if method not in methods:
            raise ValueError(f"'{method}' is not a valid clipping method. Select "
                             f"from {list(methods)}")
        retval = methods[method]
        logger.debug("[GradClip] Got clipper '%s': %s", method, retval)
        return retval

    def __call__(self, parameters: list[nn.Parameter]) -> None:
        """Clip the given parameters by the chosen method

        Parameters
        ----------
        parameters
            The parameters to clip
        """
        self._clipper(parameters, self._value)


class Optimizer:
    """Object for managing the selected Torch optimizer

View on GitHub (pinned to f530cb7508)

Solutions

1. Change the gradient clipping setting in the training config to one of: autoclip, global_norm, norm, value.
2. If unsure of the exact spelling, use the GUI's Train > Configure settings dialog, which only offers valid values.
3. Check for stale config files after upgrading Faceswap and regenerate the config.

Example fix

# before
clipgrad = auto-clip

# after
clipgrad = autoclip

valid_clip_methods = {"autoclip", "global_norm", "norm", "value"}
assert cfg_value in valid_clip_methods, f"clip method must be one of {valid_clip_methods}"

def is_valid_clip_method(method: str) -> bool:
    return method in {"autoclip", "global_norm", "norm", "value"}

When it happens

Trigger: Editing the training config file and setting the gradient clipping option to a typo'd or unsupported value (e.g. 'auto-clip', 'gradient', 'clip'), or carrying over a value from an older Faceswap version whose method names changed.

Common situations: Hand-editing config files; config written by an older version of the codebase; names with wrong case or hyphens instead of underscores.

Related errors

• '{name}' is not a valid optimizer. Select from {list(_OPTIMI
• Penalized Mask Loss has been selected but you have not chose
• You have selected the mask type '{mask_type}' but at least o
• Load weights selected, but the path '{weights_file}' does no
• 'Learn Mask' has been selected but you have not chosen a Mas

AI-assisted analysis of deepfakes/faceswap@f530cb7508 (2026-08-15). Data as JSON: /api/errors/8b1766894b2249dc. Report an issue: GitHub.