Delay¶

class continual.Delay(delay, temporal_fill='zeros', auto_shrink=False)[source]¶

Delay an input by a number of steps.

This module only introduces a delay in the continual modes, i.e. on forward_step and forward_steps. In essence it caches the input for delay steps before outputting it again.

The Delay modules is used extensively in various container modules to align delays of different computational branches. For instance, it is used to align the Residual module as shown in the example below.

Parameters:

delay (int) – The number of steps to delay an output.
temporal_fill (PaddingMode) – Temporal state initialisation mode (“zeros” or “replicate”)
auto_shrink (Union[bool, str]) – Whether to shrink the temporal dimension of the feature map during forward. This module is handy for residuals that are parallel to modules which reduce the number of temporal steps. Options: “centered” or True: Centered residual shrink; “lagging”: lagging shrink.

Examples:

conv = co.Conv3d(32, 32, kernel_size=3, padding=1)
residual = co.BroadcastReduce(conv, co.Delay(2), reduce="sum")