deezer/spleeter · error · ValueError

Unkwnown loss type: {loss_type}

Error message

Unkwnown loss type: {loss_type}

What it means

_build_loss switches on loss_type (from model params, e.g. 'L1' or 'L1_wmag') and raises ValueError for any unrecognized value after the if/elif chain. The typo 'Unkwnown' identifies this exact branch. Only the loss types implemented in the branch chain are supported.

Source

Thrown at spleeter/model/__init__.py:213

                Tensorflow (loss, metrics) tuple.
        """
        output_dict = self.model_outputs
        loss_type = self._params.get("loss_type", self.L1_MASK)
        if loss_type == self.L1_MASK:
            losses = {
                name: tf.reduce_mean(tf.abs(output - labels[name]))
                for name, output in output_dict.items()
            }
        elif loss_type == self.WEIGHTED_L1_MASK:
            losses = {
                name: tf.reduce_mean(
                    tf.reduce_mean(labels[name], axis=[1, 2, 3], keep_dims=True)
                    * tf.abs(output - labels[name])
                )
                for name, output in output_dict.items()
            }
        else:
            raise ValueError(f"Unkwnown loss type: {loss_type}")
        loss = tf.reduce_sum(list(losses.values()))
        # Add metrics for monitoring each instrument.
        metrics = {k: tf.compat.v1.metrics.mean(v) for k, v in losses.items()}
        metrics["absolute_difference"] = tf.compat.v1.metrics.mean(loss)
        return loss, metrics

    def _build_optimizer(self) -> tf.Tensor:
        """
        Builds an optimizer instance from internal parameter values.
        Default to AdamOptimizer if not specified.

        Returns:
            tf.Tensor:
                Optimizer instance from internal configuration.
        """
        name = self._params.get("optimizer")
        if name == self.ADADELTA:
            return tf.compat.v1.train.AdadeltaOptimizer()

View on GitHub (pinned to c8854001ac)

Solutions

1. Set loss_type to a supported value, e.g. 'L1' or 'L1_wmag', matching the exact spelling/case in the source
2. Print/inspect the code in _build_loss to see the accepted branch values
3. Remove the loss_type override so the default path is used
4. Implement a custom branch in _build_loss if you truly need another loss

Example fix

// before (config)
{"params": {"loss_type": "MSE"}}
// after
{"params": {"loss_type": "L1"}}

SUPPORTED_LOSSES = {'L1', 'L1_wmag'}  # matches _build_loss branches
loss_type = params.get('loss_type')
if loss_type is not None and loss_type not in SUPPORTED_LOSSES:
    raise ValueError(f"Unsupported loss_type {loss_type!r}; supported: {SUPPORTED_LOSSES}")

def is_supported_loss(cfg: dict) -> bool:
    return cfg.get('model', {}).get('params', {}).get('loss_type', 'L1') in {'L1', 'L1_wmag'}

try:
    builder.build_train_model(labels)
except ValueError as e:
    if 'Unkwnown loss type' in str(e):
        raise ConfigError(f"Invalid loss_type in model params: {e}") from e
    raise

When it happens

Trigger: Setting model.params.loss_type (or the equivalent config field) to something other than the supported values, e.g. 'l2', 'mse', 'L1 ', 'l1' (case mismatch), or a custom loss not implemented in this spleeter version.

Common situations: Copying a config from another separation library; assuming case-insensitive matching; upgrading/downgrading spleeter so a previously supported loss type no longer exists.

Related errors

• No model function {model_type} found
• Invalid mask_extension parameter {extension}
• n_chunks_per_song must be positif
• Unknown mode {mode}
• {adapter_class_name} is not a valid AudioAdapter class

AI-assisted analysis of deezer/spleeter@c8854001ac (2026-08-28). Data as JSON: /api/errors/504edf8929610d70. Report an issue: GitHub.