Skip to content

TODO.rst

Latest commit

 

History

History
105 lines (90 loc) · 3.54 KB

TODO.rst

File metadata and controls

105 lines (90 loc) · 3.54 KB
  • cite for raggedarray: Awkward Array, Cython, Numba -> Check it out. AA uses cython. Doesn't support some things
  • cite/mention biopython for io NCBI retrieval -> How to combine bionumpiy with biopython
  • explain raggedarray implementation (second paragraph). Performance cost -> benchmark, indexing, reductions, elementwise operations, accumulations: see asv
    • asv Ivar lead (1 day maybe pair)
  • Pip->pip (typo)

Third paragraph

  • Mention supported data file formats/ fasta and genbank files, maybe mention biopython -> What is genbank files (maybe support)

Fourth paragraph

  • "In" before "Figure 1" (typo)
  • axis labels in figure 1 (typo)

General

  • array API: https://data-apis.org/array-api/2022.12/API_specification/index.html (1 day Knut)
    • Two points. Support as much as possible. Support in backend as much as possible (changing backend)
  • interplay with torch, jax, numpy -> (1 day pair)
    • change backend to jax (maybe torch), mention cupy.
    • Make work with test suite
    • Hacks in npstructures not in bionumpy
  • Autodiff -> (implied)
    • If jax works everything works
  • Data types. string/unicode numeric SI?
    -- ACII for bulk data -- Encoded arrays for DNA, RNA -- Maybe accept unicode

Benchmarks - array API: https://data-apis.org/array-api/2022.12/API_specification/index.html

  • from importlib.metadata import entry_points
  • return hasattr(x, '__array_namespace__')
  • https://data-apis.org/array-api/2022.12/verification_test_suite.html
  • doesn't support list inputs
  • mutation is not allowed
  • access device in raggedarrays
  • dlpack
  • .devce, device=None, to_device
  • in-place operations?
  • we can't do T, mT
  • __array_namespace__
  • __index__
  • __int__
  • broadcasting - we only support 2 dims
  • array creation algorithms
  • positional only?
  • must return zero-dim array on indexing (this differs from numpy)
  • boolean indexing is voluntary
  • take
  • manipulation functions
  • argmin, argmax, nonzero, where
  • set functions
  • argsort, sort
  • max, mean, min, prod, std, sum, var, all, any
  • check type promotion, should be delegated entirely to numpy
  • __array_api_version__
  • extension name
  • interplay with torch jax, numpy
  • Autodiff
  • Data types. string/unicode numeric SI?- Supporting Info
  • Benchmarks
  • Use asv: https://asv.readthedocs.io/en/v0.6.1/) - Use in npstructures - Use also a bit in bionumpy
  • y-axis label on each subplot
  • Repeats for benchmarks, include variance
    • Make multiple repetitions, and do variance
    • Explain that we benchmark against non-python code
  • Make enformer assesment reproducible (if possible)
  • "PiP" -> "pip"

Implementation Details

  • Figure 2: 2 x 8 = 24 bytes?
  • explain vectorization. reduce-at - ties in to performance cost earlier
  • ufuncs on encoded counts
-- Maybe implement, low priority
  • master -> main

Support alignment - Data structures [MFA, MAF, CLUSTAL] -> Should be easy - Algorithms [Wrapper for bwa-mem, minimap2, clustalw, clustalo] -> Only if it's relatively easy (1 day)

mapper = bwa_mem('hg38.fa') aligned_reads = mapper.align('reads.fq')
  • Improve documentation -> Timebox (1/2 day together)
    • Add docstrings for all public things
    • Maybe fix typing + add card for it
    • Go a bit through the main documentation and clean up
    • Have 4 concrete stories that should be possible to do with bionumpy
      • interplay with jax and so on
      • end to end pipeline with wrappers
      • immune classification
      • kage
      • (deeptools example)
      • signal plot and matrix
  • Blog post -> Chakri (1 day)