Reduce Import Time

[seanlaw]

The time that it takes for STUMPY to get imported has been growing. This is assessed via:

$ python -X importtime -c "import stumpy"

Which shows the bulk of the import time being spent here:

...
import time:   1476386 |    1820531 |   stumpy.core
import time:   1137810 |    1137810 |     stumpy.aamp
import time:   2498991 |    3636800 |   stumpy.stump
import time:       342 |        342 |     stumpy.aamped
import time:       476 |        817 |   stumpy.stumped
import time:    844657 |     844657 |     stumpy.maamp
import time:    718504 |    1563161 |   stumpy.mstump
import time:       351 |        351 |     stumpy.maamped
import time:       472 |        823 |   stumpy.mstumped
import time:       301 |        301 |   stumpy.aampi
import time:       195 |        195 |   stumpy.chains
import time:       205 |        205 |   stumpy.floss
import time:       254 |        254 |     stumpy.aamp_ostinato
import time:       291 |        544 |   stumpy.ostinato
import time:   1209154 |    1209154 |     stumpy.scraamp
import time:    794046 |    2003200 |   stumpy.scrump
import time:      1048 |       1048 |   stumpy.stumpi
import time:       160 |        160 |     stumpy.aampdist
import time:       199 |        358 |   stumpy.mpdist
import time:       290 |        290 |     stumpy.aamp_motifs
import time:       183 |        473 |   stumpy.motifs
import time:       146 |        146 |     stumpy.aampdist_snippets
import time:       174 |        320 |   stumpy.snippets
import time:       236 |        236 |   stumpy.stimp
import time:      6004 |    9125575 | stumpy

It would be great to figure out how to reduce this import time down. It appears that it has something to do with how we are importing functions into __init__.py

[seanlaw]

https://numba.discourse.group/t/reducing-import-time/1072/10

[seanlaw]

When we do import stumpy, these lines of code force Python to import ALL functions:

stumpy/stumpy/__init__.py

Lines 9 to 70 in 27e1eed

	from . import cache, config
	from .aamp import aamp # noqa: F401
	from .aamp_mmotifs import aamp_mmotifs # noqa: F401
	from .aamp_motifs import aamp_match, aamp_motifs # noqa: F401
	from .aamp_ostinato import aamp_ostinato, aamp_ostinatoed # noqa: F401
	from .aamp_stimp import aamp_stimp, aamp_stimped # noqa: F401
	from .aampdist import aampdist, aampdisted # noqa: F401
	from .aampdist_snippets import aampdist_snippets # noqa: F401
	from .aamped import aamped # noqa: F401
	from .aampi import aampi # noqa: F401
	from .chains import allc, atsc # noqa: F401
	from .core import mass # noqa: F401
	from .floss import floss, fluss # noqa: F401
	from .maamp import maamp, maamp_mdl, maamp_subspace # noqa: F401
	from .maamped import maamped # noqa: F401
	from .mmotifs import mmotifs # noqa: F401
	from .motifs import match, motifs # noqa: F401
	from .mpdist import mpdist, mpdisted # noqa: F401
	from .mstump import mdl, mstump, subspace # noqa: F401
	from .mstumped import mstumped # noqa: F401
	from .ostinato import ostinato, ostinatoed # noqa: F401
	from .scraamp import prescraamp, scraamp # noqa: F401
	from .scrump import prescrump, scrump # noqa: F401
	from .snippets import snippets # noqa: F401
	from .stimp import stimp, stimped # noqa: F401
	from .stump import stump # noqa: F401
	from .stumped import stumped # noqa: F401
	from .stumpi import stumpi # noqa: F401
	# Get the default fastmath flags for all njit functions
	# and update the _STUMPY_DEFAULTS dictionary
	if not numba.config.DISABLE_JIT: # pragma: no cover
	njit_funcs = cache.get_njit_funcs()
	for module_name, func_name in njit_funcs:
	module = importlib.import_module(f".{module_name}", package="stumpy")
	func = getattr(module, func_name)
	key = module_name + "." + func_name # e.g., core._mass
	key = "STUMPY_FASTMATH_" + key.upper() # e.g., STUMPY_FASTHMATH_CORE._MASS
	config._STUMPY_DEFAULTS[key] = func.targetoptions["fastmath"]
	if cuda.is_available():
	from .gpu_aamp import gpu_aamp # noqa: F401
	from .gpu_aamp_ostinato import gpu_aamp_ostinato # noqa: F401
	from .gpu_aamp_stimp import gpu_aamp_stimp # noqa: F401
	from .gpu_aampdist import gpu_aampdist # noqa: F401
	from .gpu_mpdist import gpu_mpdist # noqa: F401
	from .gpu_ostinato import gpu_ostinato # noqa: F401
	from .gpu_stimp import gpu_stimp # noqa: F401
	from .gpu_stump import gpu_stump # noqa: F401
	else: # pragma: no cover
	from . import core
	from .core import _gpu_aamp_driver_not_found as gpu_aamp # noqa: F401
	from .core import ( # noqa: F401
	_gpu_aamp_ostinato_driver_not_found as gpu_aamp_ostinato,
	)
	from .core import _gpu_aamp_stimp_driver_not_found as gpu_aamp_stimp # noqa: F401
	from .core import _gpu_aampdist_driver_not_found as gpu_aampdist # noqa: F401
	from .core import _gpu_mpdist_driver_not_found as gpu_mpdist # noqa: F401
	from .core import _gpu_ostinato_driver_not_found as gpu_ostinato # noqa: F401
	from .core import _gpu_stimp_driver_not_found as gpu_stimp # noqa: F401
	from .core import _gpu_stump_driver_not_found as gpu_stump # noqa: F401

Note that the fastmath default flags are particularly problematic and further increase the import time because it imports all njit functions. Instead, we can possibly implement a "lazy import" approach that only imports the functions as needed.