Pluggable point-loss functions for the DLinear forecaster (self-contained). Each loss has signature (pred, target, mask=None) -> scalar. When mask is given it is a per-element 0/1 weight and the reduction is the masked mean sum(loss*mask)/sum(mask) (denominator clipped to >= 1, which equals NF's divide-no-nan on an all-masked batch) — matching neuralforecast's _weighted_mean, used to drop right-padded horizon steps from the training loss.
mae
dlinear_losses.mae
Mean absolute error (masked mean when mask is given).
| Parameter | Type | Default | Description |
|---|---|---|---|
pred |
jnp.ndarray |
- | (undocumented) |
target |
jnp.ndarray |
- | (undocumented) |
mask |
jnp.ndarray | None |
None |
(undocumented) |
Returns: jnp.ndarray
mse
dlinear_losses.mse
Mean squared error (masked mean when mask is given).
| Parameter | Type | Default | Description |
|---|---|---|---|
pred |
jnp.ndarray |
- | (undocumented) |
target |
jnp.ndarray |
- | (undocumented) |
mask |
jnp.ndarray | None |
None |
(undocumented) |
Returns: jnp.ndarray
huber
dlinear_losses.huber
Huber loss with delta = 1.0 (masked mean when mask is given).
| Parameter | Type | Default | Description |
|---|---|---|---|
pred |
jnp.ndarray |
- | (undocumented) |
target |
jnp.ndarray |
- | (undocumented) |
mask |
jnp.ndarray | None |
None |
(undocumented) |
Returns: jnp.ndarray
LOSSES
dlinear_losses.LOSSES
Mapping of available loss function names to their implementations.
Type: Mapping[str, LossFn]
resolve
dlinear_losses.resolve
Return a callable loss from either a registry string or a callable.
Callables receive (pred, target, mask); a custom loss should accept an optional mask (def my_loss(pred, target, mask=None): ...).
| Parameter | Type | Default | Description |
|---|---|---|---|
loss |
str | LossFn |
- | (undocumented) |
Returns: LossFn
Raises:
* ValueError