perf: don't apply fancy indexing if split can be a slice.

[selmanozleyen]

In my use case when batch_size=1, splits becomes something like this = [[0],[1],[2],...]. For each split I think this becomes a fancy indexing operation and can be costly. In my case every row is 40mb. And I want to have some preloaded rows but turns out each in_memory_data[split] is costly even though it fetches for one row.

This is a bit more of a generalized case. But I added (split[-1] - split[0] == len(split) - 1 check so it returns early but if you think we should avoid this check we can also just have a special case for batch_size=1

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.92%. Comparing base (796e789) to head (d520726).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #235      +/-   ##
==========================================
- Coverage   93.48%   91.92%   -1.57%     
==========================================
  Files          15       15              
  Lines        1397     1399       +2     
==========================================
- Hits         1306     1286      -20     
- Misses         91      113      +22     

Files with missing lines	Coverage Δ
src/annbatch/loader.py	87.79% <100.00%> (-3.01%)	⬇️

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ilan-gold]

In my use case when batch_size=1

Just to be clear, then

sel = slice(sel[0], sel[-1] + 1)

becomes a length-1 slice? And this is faster than fancy-indexing? Feels like a bug almost (not with us)

[selmanozleyen]

In my use case when batch_size=1

Just to be clear, then

sel = slice(sel[0], sel[-1] + 1)

becomes a length-1 slice? And this is faster than fancy-indexing? Feels like a bug almost (not with us)

Yes its setup: preloadnchunks=x,batchsize=1,chunksize1.
It is because I have irregular batch sizes but I also want the preloadnchunks of annbatch. Like unique donors corresponding to a batch can be arbitrary. I asked Lukas if its fine if we sample fixed sized donors and sample the corresponding cells from the donors so we have fixed sized batches and it would be somewhat mathematically equivalent. But I wanted to implement whatever lucas already has 1-1

[selmanozleyen]

I checked and for numpy it creates a copy in case of fany indexing of the size of the index. Vs compared to slices which only returns a view. Since each element is 40mb even allocating one row can be costly. In short if its a slice its a view if its a list its a copy view vs copy. Makes sense that they don't have special cases because you would want consistent behaviour.