dymad.models.prediction

Functions

predict_continuous(model, x0, ts[, ws, ...])

Predict trajectory(ies) for continuous-time models with batch support.

predict_continuous_exp(model, x0, ts[, ws])

Predict trajectory(ies) for continuous-time models with batch support.

predict_continuous_fenc(model, x0, ts[, ws])

Predict trajectory(ies) using First-order Euler with Normal Correction (FENC).

predict_continuous_np(model, x0, ts[, ws, ...])

Predict trajectory(ies) for continuous-time models with batch support.

predict_discrete(model, x0, ts[, ws])

Predict trajectory(ies) for discrete-time models with batch support.

predict_discrete_exp(model, x0, ts[, ws])

Predict trajectory(ies) for discrete-time models with batch support.
dymad.models.prediction.predict_continuous(model, x0, ts, ws=None, method='dopri5', order='cubic', **kwargs)

Predict trajectory(ies) for continuous-time models with batch support.

Parameters:

  • model – Model with encoder, decoder, and dynamics methods.

  • x0 (torch.Tensor) –

    Initial state(s).

    • Single: shape (n_features,)

    • Batch: shape (batch_size, n_features)

  • ts (Union[np.ndarray, torch.Tensor]) – Time points (n_steps,).

  • ws (TypeAliasType | None) – Dataclass containing additional information, e.g., u, p, ei, ew, etc.

  • method (str) – ODE solver method (default: ‘dopri5’).

  • order (str) – Interpolation method for control inputs (‘zoh’, ‘linear’, or ‘cubic’).

Returns:

Predicted trajectory(ies).

  • Single: shape (n_steps, n_features)

  • Batch: shape (batch_size, n_steps, n_features)

Return type:

torch.Tensor

dymad.models.prediction.predict_continuous_exp(model, x0, ts, ws=None, **kwargs)

Predict trajectory(ies) for continuous-time models with batch support.

Autonomous case using matrix exponential. In continuous-time, we compute exp(A*dt).

The step size is assumed to be constant, so when there are batched time series, we just take the first one to compute dt.

Currently only for KBF-type models with linear dynamics.

Return type:

Tensor

dymad.models.prediction.predict_continuous_fenc(model, x0, ts, ws=None, **kwargs)

Predict trajectory(ies) using First-order Euler with Normal Correction (FENC).

Currently only for kernel machine with tangent kernel.

Return type:

Tensor

dymad.models.prediction.predict_continuous_np(model, x0, ts, ws=None, method='dopri5', order='cubic', **kwargs)

Predict trajectory(ies) for continuous-time models with batch support.

No-projection version, meaning during ODE integration, we do not decode back to the observation space and encode back; the decoding happens only at the end.

Return type:

Tensor

dymad.models.prediction.predict_discrete(model, x0, ts, ws=None, **kwargs)

Predict trajectory(ies) for discrete-time models with batch support.

Parameters:

  • model – Model with encoder, decoder, and dynamics methods

  • x0 (Tensor) –

    Initial state(s):

    • Single: (n_features,)

    • Batch: (batch_size, n_features)

  • ts (Union[np.ndarray, torch.Tensor]) – Time points (n_steps,).

  • ws (TypeAliasType | None) – Dataclass containing additional information, e.g., u, p, ei, ew, etc.

Returns:

Predicted trajectory(ies)

  • Single: (n_steps, n_features)

  • Batch: (batch_size, n_steps, n_features)

Return type:

torch.Tensor

dymad.models.prediction.predict_discrete_exp(model, x0, ts, ws=None, **kwargs)

Predict trajectory(ies) for discrete-time models with batch support.

In discrete-time, this is equivalent to repeated application of the dynamics.

Return type:

Tensor