Stability-AI/generative-models · warning

Did not find parameters for pattern {pattern_}

Error message

Did not find parameters for pattern {pattern_}

What it means

get_param_groups builds optimizer parameter groups from regex patterns over named_parameters. When a pattern matches zero parameters it logs a warning 'Did not find parameters for pattern ...' and continues with an empty group — the user likely misspelled a pattern, so part of the model silently gets no (or wrong) training configuration.

Source

Thrown at sgm/models/autoencoder.py:358

        self.log_dict(full_log_dict, sync_dist=True)
        return full_log_dict

    def get_param_groups(
        self, parameter_names: List[List[str]], optimizer_args: List[dict]
    ) -> Tuple[List[Dict[str, Any]], int]:
        groups = []
        num_params = 0
        for names, args in zip(parameter_names, optimizer_args):
            params = []
            for pattern_ in names:
                pattern_params = []
                pattern = re.compile(pattern_)
                for p_name, param in self.named_parameters():
                    if re.match(pattern, p_name):
                        pattern_params.append(param)
                        num_params += param.numel()
                if len(pattern_params) == 0:
                    logpy.warn(f"Did not find parameters for pattern {pattern_}")
                params.extend(pattern_params)
            groups.append({"params": params, **args})
        return groups, num_params

    def configure_optimizers(self) -> List[torch.optim.Optimizer]:
        if self.trainable_ae_params is None:
            ae_params = self.get_autoencoder_params()
        else:
            ae_params, num_ae_params = self.get_param_groups(
                self.trainable_ae_params, self.ae_optimizer_args
            )
            logpy.info(f"Number of trainable autoencoder parameters: {num_ae_params:,}")
        if self.trainable_disc_params is None:
            disc_params = self.get_discriminator_params()
        else:
            disc_params, num_disc_params = self.get_param_groups(
                self.trainable_disc_params, self.disc_optimizer_args
            )

View on GitHub (pinned to e8cd657656)

Solutions

1. Print [n for n, _ in model.named_parameters()] and fix the regex patterns to match actual names
2. Verify trainable_ae_params/trainable_disc_params entries in the config match the model being instantiated
3. Treat the warning as fatal during development (raise or assert) to catch config mistakes early

Example fix

// before
trainable_ae_params: [[{"name": "decder", "pattern": "decder\.", "lr": 1e-4}]]  # typo
// after
trainable_ae_params: [[{"name": "decoder", "pattern": "^decoder\.", "lr": 1e-4}]]

import re
for group in trainable_params:
    matched = [n for n, _ in model.named_parameters() if re.search(group["pattern"], n)]
    assert matched, f"pattern {group['pattern']} matches nothing"

def pattern_matches(model, pattern: str) -> bool:
    return any(re.search(pattern, n) for n, _ in model.named_parameters())

try:
    groups, n = model.get_param_groups()
except Exception:
    for name, _ in model.named_parameters():
        print(name)  # debug actual names

When it happens

Trigger: Configuring trainable_ae_params / trainable_disc_params with regex patterns like '^encoder\.' that do not match any parameter names of the autoencoder (e.g. pattern 'decoder.conv_in' when names are prefixed differently, or using '.*disc.*' on a model without a discriminator).

Common situations: Copy-pasting optimizer configs between AutoencoderKL variants, typos in YAML regex patterns, switching models where parameter names changed, or freezing everything so the pattern's params are excluded from named_parameters.

Related errors

• unknown merge strategy {self.merge_strategy}
• Unknown loss type {self.loss_type}
• provide num_res_blocks either as an int (globally constant)
• need either 'input_key' or 'input_keys' for embedder {embedd
• NotImplementedError

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/451be0ba8a56394f. Report an issue: GitHub.