Feature request: faster import

Is your feature request related to a problem? Please describe.

It’s very slow to import sfaira. This includes any use of the command line interface. For example:

$ time sfaira --help
...
sfaira --help  28.27s user 1.04s system 106% cpu 27.624 total

%time import sfaira

Describe the solution you’d like

Faster import of Sfaira by delaying computation of whatever is taking so long.

Additional context

It looks to me like this is mostly coming from this line:

https://github.com/theislab/sfaira/blob/dab9cb3f6d89dc6c6b52f9ce23c9a7e938e4d91b/sfaira/consts/__init__.py#L5

and the subsequent loading of a bunch of ontologies. Here’s the first couple lines of %prun -s cumtime import sfaira for evidence that this is the culprit:

         46387199 function calls (46339180 primitive calls) in 33.690 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
      276    0.005    0.000   38.698    0.140 __init__.py:1(<module>)
        7    0.001    0.000   33.802    4.829 __init__.py:2(<module>)
   4477/1    0.029    0.000   33.693   33.693 {built-in method builtins.exec}
   3377/1    0.014    0.000   33.693   33.693 <frozen importlib._bootstrap>:986(_find_and_load)
   3364/1    0.010    0.000   33.693   33.693 <frozen importlib._bootstrap>:956(_find_and_load_unlocked)
   3166/1    0.011    0.000   33.693   33.693 <frozen importlib._bootstrap>:650(_load_unlocked)
   2975/1    0.006    0.000   33.693   33.693 <frozen importlib._bootstrap_external>:842(exec_module)
   5107/1    0.003    0.000   33.693   33.693 <frozen importlib._bootstrap>:211(_call_with_frames_removed)
        1    0.000    0.000   30.572   30.572 ontologies.py:10(__init__)
        6    0.000    0.000   29.722    4.954 base.py:551(__init__)
        6    0.045    0.008   29.722    4.954 base.py:512(__init__)
        6    0.657    0.109   29.677    4.946 read.py:12(read_obo)
        6    1.020    0.170   28.010    4.668 read.py:63(get_sections)

[davidsebfischer]

We could keep the OCS constant but only populate it with the intialised ontologies based on a particular function call, sfaira.initialise() for example.

[Zethson]

https://github.com/theislab/sfaira/commit/c26c0dbbbcf07cc82169a960a70ea6f88595af64

@davidsebfischer CLI startup is now in the order of 1-2 seconds. Can’t improve it much further as long as TensorFlow is involved.