DLinearNet

dlinear_module.DLinearNet · inherits nnx.Module

forecast = trend @ Wt.T + bt + seasonal @ Ws.T + bs (no last-value add-back — that is NLinear's trick). I/O [B, L, 1] -> [B, h, 1].

__init__(self, h: int, input_size: int, moving_avg_window: int, *, rngs: nnx.Rngs)

Initializes the DLinear network parameters (weights and biases for both trend and seasonal linear heads). Weights are initialized uniformly based on the input size, replicating the distribution used in NeuralForecast/PyTorch.

Parameter	Type	Default	Description
h	int	-	(undocumented)
input_size	int	-	(undocumented)
moving_avg_window	int	-	(undocumented)
rngs	nnx.Rngs	-	(undocumented)

__call__(self, x: jnp.ndarray) -> jnp.ndarray

Performs series decomposition on the input, applies separate linear transformations to the trend and seasonal components, and sums the results to produce the forecast.

Parameter	Type	Default	Description
x	jnp.ndarray	-	(undocumented)

Returns: jnp.ndarray (Output tensor, shape [B, h, 1]).