DeepAR_EncDec

chronax.model.DeepAR_EncDec · inherits nn.Module

(No prose description provided in class docstring.)

__init__(self, hidden=64, dropout_rate=0.1, min_sigma=0.02)

Parameter	Type	Default	Description
hidden	int	64	(undocumented)
dropout_rate	float	0.1	(undocumented)
min_sigma	float	0.02	(undocumented)

encode(self, y_hist, futr_exog=None, x_static=None)

(No prose summary.)

Parameter	Type	Default	Description
y_hist	-	-	(undocumented)
futr_exog	-	None	(undocumented)
x_static	-	None	(undocumented)

Returns: hT, cT

one_step(self, y_prev, x_f_step=None, x_static=None, h=None, c=None, deterministic=True)

(No prose summary.)

Parameter	Type	Default	Description
y_prev	-	-	(undocumented)
x_f_step	-	None	(undocumented)
x_static	-	None	(undocumented)
h	-	None	(undocumented)
c	-	None	(undocumented)
deterministic	bool	True	(undocumented)

Returns: mu[0], sigma[0], h_new, c_new

__call__(self, y_seq, futr_exog=None, x_static=None, training=True, init_state=None)

(No prose summary.)

Parameter	Type	Default	Description
y_seq	-	-	(undocumented)
futr_exog	-	None	(undocumented)
x_static	-	None	(undocumented)
training	bool	True	(undocumented)
init_state	-	None	(undocumented)

Returns: mu, sigma

make_trainer

chronax.model.make_trainer

(No prose summary.)

Parameter	Type	Default	Description
model	-	-	(undocumented)
lr	-	1e-3	(undocumented)
weight_decay	-	1e-5	(undocumented)

Returns: tx, step

train_model

chronax.model.train_model

Train DeepAR with batched encode-decode and seasonal lag features.

Lag features: lag-1 (autocorr), lag-7 (weekly), lag-h (horizon), lag-2h. During training, all lags are precomputed from teacher-forced windows. During inference, lag-1/7 are tracked via a rolling buffer in the scan carry.

Parameter	Type	Default	Description
y_hist	jnp.ndarray	-	(undocumented)
x_f_all	jnp.ndarray	None	(undocumented)
x_static	jnp.ndarray	None	(undocumented)
hidden	int	64	(undocumented)
lr	float	1e-3	(undocumented)
steps	int	800	(undocumented)
dropout	float	0.1	(undocumented)
min_sigma	float	0.02	(undocumented)
h	int	24	(undocumented)
batch_size	int	32	(undocumented)
verbose	bool	True	(undocumented)

Returns: (model, params, losses, scaler) where scaler = (mean, std).

forecast_mc

chronax.model.forecast_mc

Monte Carlo forecast. Returns sample paths [N, H].

Lag-1/7 are tracked via a rolling buffer in the scan carry so predicted values are correctly used as lags in later decode steps.

Parameter	Type	Default	Description
params	-	-	(undocumented)
model	DeepAR_EncDec	-	(undocumented)
y_hist	jnp.ndarray	-	(undocumented)
x_f_hist	jnp.ndarray	None	(undocumented)
x_f_future	jnp.ndarray	None	(undocumented)
x_static	jnp.ndarray	None	(undocumented)
H	int	24	(undocumented)
N	int	1000	(undocumented)
seed	int	2025	(undocumented)
scaler	-	None	(undocumented)
input_size	int	None	(undocumented)

Returns: jnp.ndarray (sample paths [N, H]).

quantiles

chronax.model.quantiles

(No prose summary.)

Parameter	Type	Default	Description
paths	jnp.ndarray	-	(undocumented)
qs	-	(0.1, 0.5, 0.9)	(undocumented)

Returns: list of jnp.ndarray