Memory efficient implementation for the `update()` in `OutputLagFeatureProcessor`

[garyfanhku]

Hi Jinyu,

Regarding the update in OutputLagFeatureProcessor, perhaps a little off-topic, but I wonder what would be more memory efficient other than deque here?

Thanks!

[jxx123]

Hi,

My memory is a bit vague now. I originally thought a better way would be keeping track of the lags without duplicating the original data. We only need to generate the final feature matrix in the end. For example, what we have in the beginning is just a vector, [x1, x2, x3]. Once we add lag feature, we have a matrix [[x1, x2, x3], [None, x1, x2], [None, None, x1]]. What we actually need to construct this matrix is just the vector [x1, x2, x3], and the lags, 0, 1, 2. By doing this, we can reduce the memory footprint, but it might increase the computation time, since you need to reconstruct the matrix in the end, and in the end you will still need this size of memory to store the matrix. This is a trade-off. These lag processors were implemented in a mess. Any refactoring work is welcome. Feel free to send me PRs :).