EMAReducer

class EMAReducer(step_time: float, alpha: float, duration: float = 0.0, inclusive: bool = False, inplace: bool = False)

Bases: FoldReducer

Stores the exponential moving average.

\[\begin{split}\begin{align*} s_0 &= x_0 \\ s_{t + 1} &= \alpha x_{t + 1} + (1 - \alpha) s_t \end{align*}\end{split}\]

For the smoothed data (state) \(s\) and observation \(x\) and where smoothing factor \(\alpha\) is as follows.

\[\alpha = 1 - \exp \left(\frac{-\Delta t}{\tau}\right)\]

For some time constant \(\tau\).

Parameters:

• step_time (float) – length of time between observations, \(\Delta t\).

• alpha (float) – exponential smoothing factor, \(\alpha\).

• duration (float, optional) – length of time over which results should be stored, in the same units as \(\Delta t\). Defaults to 0.0.

• inclusive (bool, optional) – if the duration should be inclusive. Defaults to False.

• inplace (bool, optional) – if write operations should be performed in-place. Defaults to False.

Note

alpha is decoupled from the step time, so if the step time changes, then the

underlying time constant will change, alpha will remain the same.

fold(obs: Tensor, state: Tensor | None) → Tensor

Application of exponential smoothing.

Parameters:

• obs (torch.Tensor) – observation to incorporate into state.

• state (torch.Tensor | None) – state from the prior time step, None if no prior observations.

Returns:

state for the current time step.

Return type:

torch.Tensor

interpolate(prev_data: Tensor, next_data: Tensor, sample_at: Tensor, step_time: float) → Tensor

Linear interpolation between observations.

Parameters:

• prev_data (torch.Tensor) – most recent observation prior to sample time.

• next_data (torch.Tensor) – most recent observation subsequent to sample time.

• sample_at (torch.Tensor) – relative time at which to sample data.

• step_time (float) – length of time between the prior and subsequent observations.

Returns:

interpolated data at sample time.

Return type:

torch.Tensor