Skip to content

Latest commit

 

History

History
70 lines (57 loc) · 2.71 KB

README.md

File metadata and controls

70 lines (57 loc) · 2.71 KB
Raw

NBLAST for large datasets

Experimental version of navis' NBLAST containing a few tweaks to enable blasting of very large (upwards 100k) sets of neurons.

❗ If you need to NBLAST only a few thousand neurons, you are probably better off with the implementation in navis!

At this point we offer two variants:

1. NBLAST to disk

This implementation immediately writes scores to disk. Notably, we:

  1. Use Zarr to create an on-disk array which is then populated in parallel by the spawned NBLAST processes.
  2. Use Dask to wrap the Zarr array as on-disk DataFrame.
  3. Some (minimal) effort to balance the load for each process.
>>> from nblast import nblast_disk
>>> # `queries`/`targets` are navis.Dotprops
>>> scores = .nblast_disk(queries, targets,
...                       out='./nblast_results'  # where to store Zarr array with results
...                       return_frame=True)      # default is `return_frame=False`
>>> # Convert Dask to in-memory pandas DataFrame
>>> scores = scores.compute()

The scores are stored as Zarr array in ./nblast_results/scores. To re-open:

>>> import zarr
>>> # Open the array
>>> z = zarr.open_array("./nblast_results/scores", mode="r")
>>> # IDs of queries/targets are stored as attributes
>>> z.attrs['queries']
[5812980561, 5813056261, 1846312261, 2373845661, ...]

Why Zarr for the on-disk array?

Because Zarr has a very easy to use interface for coordinating parallel writes.

Is this slower than the original navis implementation?

Not as far as I can tell: the overhead from using Zarr/Dask appears surprisingly low. It does, however, introduce new dependencies and the returned Dask DataFrame is more difficult to handle than run-of-the-mill pandas DataFrames.

2. NBLAST to sparse matrix

This implementation drops scores below a given threshold and stores results in a sparse matrix. This is useful if you don't care about low (say less than 0) scores.

>>> from nblast import nblast_sparse
>>> # `queries`/`targets` are navis.Dotprops
>>> scores = nblast_sparse(queries, targets, threshold=0)
>>> scores.dtypes
603270961     Sparse[float32, nan]
2312333561    Sparse[float32, nan]
1934410061    Sparse[float32, nan]
...
>>> scores.sparse.density
0.05122

scores is a pandas DataFrame where each column is a sparse array. It will, in many ways, handle like a normal DataFrame. See the docs for details. As things stand, you will have to convert the DataFrame dense (scores.sparse.to_dense()) in order to save it to disk (e.q. as feather or parquet) unless you want to use pickle.