microsoft/qlib · error · ValueError

unknown loss `%s`

Error message

unknown loss `%s`

What it means

Raised by SANDWITCH model loss_fn in qlib/contrib/model/pytorch_sandwich.py:241 when self.loss is not 'mse'. The sandwich model implements exactly one supervised loss — masked MSE over non-NaN labels — so any other loss string is a ValueError. The loss attribute comes from the constructor's loss kwarg.

Source

Thrown at qlib/contrib/model/pytorch_sandwich.py:241

        self.fitted = False
        self.sandwich_model.to(self.device)

    @property
    def use_gpu(self):
        return self.device != torch.device("cpu")

    def mse(self, pred, label):
        loss = (pred - label) ** 2
        return torch.mean(loss)

    def loss_fn(self, pred, label):
        mask = ~torch.isnan(label)

        if self.loss == "mse":
            return self.mse(pred[mask], label[mask])

        raise ValueError("unknown loss `%s`" % self.loss)

    def metric_fn(self, pred, label):
        mask = torch.isfinite(label)

        if self.metric in ("", "loss"):
            return -self.loss_fn(pred[mask], label[mask])

        raise ValueError("unknown metric `%s`" % self.metric)

    def train_epoch(self, x_train, y_train):
        x_train_values = x_train.values
        y_train_values = np.squeeze(y_train.values)
        self.sandwich_model.train()

        indices = np.arange(len(x_train_values))
        np.random.shuffle(indices)

        for i in range(len(indices))[:: self.batch_size]:

View on GitHub (pinned to 79633dd950)

Solutions

  1. Set loss: 'mse' in the sandwich model kwargs (currently the only option).
  2. If another loss is required, subclass and override loss_fn() (keep the NaN mask) rather than modifying the library.

Example fix

# before
kwargs:
  loss: mae

# after
kwargs:
  loss: mse
Defensive patterns

Strategy: validation

Validate before calling

assert config.get("loss", "mse") == "mse", "SANDWITCH model only supports loss='mse'"

Type guard

def is_supported_sandwich_loss(loss: str) -> bool:
    return loss == "mse"

Try / catch

try:
    model.fit(dataset)
except ValueError as e:
    if "unknown loss" in str(e):
        raise ValueError("Set loss='mse'; other losses require overriding loss_fn()") from e
    raise

Prevention

When it happens

Trigger: Passing loss='mae', loss='huber', etc. to the sandwich model and calling fit(); the first train/valid epoch evaluation calls loss_fn and raises. Note metric in ('', 'loss') also funnels into loss_fn, so a bad loss breaks metric evaluation too.

Common situations: Reusing kwargs from DNNModelPytorch (which accepts 'binary') or from other contrib models with richer loss menus; expecting symmetric naming with LightGBM's objective strings.

Related errors

AI-assisted analysis of microsoft/qlib@79633dd950 (2026-08-15). Data as JSON: /api/errors/c23f5ad5440126b1. Report an issue: GitHub.