Temporal sampling + batching strategies #14
Description
The upweighting strategy to emphasize signal based on summary images is useful. There are some things to look at on the temporal side:
We should now revisit the following things:
(1) What's the easiest way to sample temporal subsets of a dataset for the spatial fit? Before we were using a combination of window chunks and frames.
(2) How does batching work in this context? For e.g. to do a spatial fit using 150K frames (which may be too large to fit into memory) how does the existing batching procedure handle this?
(3) What are the payoffs of fusing these steps here? We use sketching to compute per block shifts; this is just a random projection. We load data --> project --> piece things together.