Frequency of ACF calculation

[amruthesht]

Additional feature request

The current ACF PR #37 uses an internal buffer to store trajectory data and provides the ability to run lag-time based calculations every time a new frame is received/processed. However, it might be useful to provide the option to change the frequency of doing such calculations.
However, to avoid statistical correlations and artificial spikes in an FFT, one should only do them either every time a new frame arrives or only when a new complete set of frames replaces the current ones in the buffer. This would mean that for a buffer of length $N$, the calculation will only happen every $N$ steps.

[amruthesht]

@PardhavMaradani -- not high priority right now. Please do it when you have time!

[PardhavMaradani]

This would mean that for a buffer of length $N$, the calculation will only happen every $N$ steps.

As of now, the buffer size and the batch size are the same (probably a separate discussion if at all this needs to be separated in the future). So, if we support batching for this widget, the computation is invoked every time the batch is full - ie., after every N steps as you suggested. There is already parallelization support for this widget, so it will help with this heavy batch computation as well. Will add this support. Thanks

[PardhavMaradani]

Hi @amruthesht , I tried this out and ran into an issue. The values calculated after every frame did not match the values calculated after every N steps. After debugging the reason, I am not sure if this would be possible to match - could you please help?

For example, consider N = 5

[ 0 1 2 3 4 ] [ 5 6 7 8 9] [ 10 ... ]

In the every-frame case, there is a sliding window where frame 3 and 7 are in the same window and hence a lag of 7 - 3 is accounted. However, in the every N case, because they aren't in the same buffer ever, this isn't getting accounted. Am I interpreting this correctly?

Thanks