Lazy arrays

[kushalkolar]

Implements #139

Dask was not a good option, too complex for our use case, so just created a base array type.

todo:

• base LazyArray
• Add min and max properties to get the min and max of if-fully-computed array.
• RCMArray
• RCBArray
• ResidualsArray
• See how fast these are with very large datasets, use a LRU cache if necessary.
  • RCM
    • single indexing is very fast, 300+ Hz with 4 RCM movies of shape 400x250 with 600-1300 components each being indexed simultaneously
    • indexing 100 slices simultaneously for all 4 results in 7 Hz, but is probably overkill amount of indexing
    • indexing 50 slices which is a reasonable upper limit results in 15 Hz which is a reasonable upper limit for indexing

    %%timeit
    i = np.random.randint(0, rcm.shape[0] - 200)
    rcm[i: i + 50]
    rcm2[i: i + 50]
    rcm3[i: i + 50]
    rcm4[i: i + 50]

  • For mesmerize-viz we can implement a "meta lazy array" which is an object containing attributes lazy_rcm_array and lazy_rcb_array so that rcm and rcb are computed only once. This object would return (RCMArray, RCBArray, and ResidualsArray) for given input indices.
• tests
• docstrings for the arrays and the cnmf extensions

⚠️ This will break the current API for cnmf.get_rcm() etc., but we are still in beta and this is much better than the current implementation.

Implemented for reconstructed movie so far, usage:

>>> rcm = RCMArray(cnmf.estimates.A, cnmf.estimates.C, cnmf.estimates.dims)
>>> rcm.shape
(3000, 170, 170)  # [n_frames, x, y]

# can do basic int indexing, works like any other array
>>> rcm[5].shape
(170, 170)

# all the fancy indexing works
>>> rcm[3:10].shape

>>> rcm[::2].shape
(1500, 170, 170)

>>> rcm[:1000:10].shape
(100, 170, 170)

Works with upcoming fastplotlib.ImageWidget

So far works with indexing for single ints and slices, but having issues when slice_avg is used

lazy_arrays-2022-12-08_01.15.50.mp4

Need to implement for rcb & residuals too.

[kushalkolar]

For ResidualsArray since it requires subtracting the RCM and RCB from the movie, we can allow rcm_lazy_array and rcb_lazy_array as optional kwargs and put a simple @lru_cache(maxsize=5) for __getitem__ in RCMArray & RCBArray so that RCM or RCB are not re-computed if being used in a visualization together.