Improve import odl speed

[adler-j]

It feels like running import odl is becomming so slow you really notice it, and this is irritating.

A short profiling using pyinstrument gave me this modest output from import odl.

2.337 <module>  odl\__init__.py:24
|- 1.392 <module>  odl\diagnostics\__init__.py:18
|  |- 0.824 <module>  odl\diagnostics\operator.py:18
|  |  `- 0.817 <module>  odl\operator\__init__.py:18
|  |     `- 0.792 <module>  odl\operator\default_ops.py:18
|  |        `- 0.785 <module>  odl\space\__init__.py:18
|  |           |- 0.684 <module>  odl\space\entry_points.py:41
|  |           |  |- 0.486 <module>  pkg_resources\__init__.py:15
|  |           |  `- 0.173 iter_entry_points  pkg_resources\__init__.py:698
|  |           `- 0.064 <module>  odl\space\weighting.py:18
|  |              `- 0.040 <module>  scipy\linalg\__init__.py:168
|  |- 0.381 <module>  odl\diagnostics\space.py:18
|  |  `- 0.373 <module>  odl\set\__init__.py:18
|  |     `- 0.359 <module>  odl\set\sets.py:18
|  |        |- 0.150 <module>  odl\util\__init__.py:18
|  |        |  `- 0.104 <module>  odl\util\testutils.py:18
|  |        |     `- 0.095 <module>  pytest.py:4
|  |        |- 0.110 <module>  past\__init__.py:85
|  |        `- 0.085 <module>  numpy\__init__.py:106
|  `- 0.187 <module>  odl\diagnostics\examples.py:18
|     |- 0.130 install_aliases  future\standard_library\__init__.py:441
|     `- 0.056 <module>  future\standard_library\__init__.py:60
|- 0.583 <module>  odl\tomo\__init__.py:18
|  |- 0.509 <module>  odl\tomo\backends\__init__.py:18
|  |  |- 0.464 <module>  odl\tomo\backends\scikit_radon.py:18
|  |  |  `- 0.457 <module>  skimage\transform\__init__.py:1
|  |  `- 0.024 <module>  odl\tomo\backends\astra_setup.py:35
|  `- 0.046 <module>  odl\tomo\geometry\__init__.py:18
|- 0.161 <module>  odl\solvers\__init__.py:18
|  |- 0.123 <module>  odl\solvers\functional\__init__.py:18
|  |  `- 0.103 <module>  odl\solvers\functional\functional.py:19
|  |     `- 0.095 <module>  odl\solvers\nonsmooth\__init__.py:18
|  |        `- 0.070 <module>  odl\solvers\nonsmooth\proximal_operators.py:25
|  |           `- 0.063 <module>  scipy\special\__init__.py:632
|  `- 0.037 <module>  odl\solvers\smooth\__init__.py:18
|- 0.057 <module>  odl\ufunc_ops\__init__.py:18
|  `- 0.057 <module>  odl\ufunc_ops\ufunc_ops.py:18
|     `- 0.044 ufunc_class_factory  odl\ufunc_ops\ufunc_ops.py:161
|- 0.052 <module>  odl\discr\__init__.py:18
|- 0.046 <module>  odl\trafos\__init__.py:18
|  `- 0.026 <module>  odl\trafos\backends\__init__.py:18
`- 0.038 <module>  odl\phantom\__init__.py:18

Here we see that there are a few main culprits in the long loading time:

• 0.7s from entry_points
• 0.45s from scikit_radon
• 0.1s from pytest
• 0.3s from future/past
• 0.1s from scipy
• 0.1s from numpy

Of these, I see that we should be able to fix at least the first 3, which should bring startup down to 0.5s which I think is more acceptable.

[kohr-h]

It feels like running import odl is becomming so slow you really notice it, and this is irritating.

Totally agree. Just today I did a much simpler exercise and compared full import times with other packages. Dropping out of Spyder and into IPython on the console and using this function

from time import time
import importlib

def test_import(pkg):
    t_start = time()
    importlib.import_module(pkg)
    t_end = time() - t_start
    print('Imported {} in {} seconds'.format(pkg, t_end))

we have

• test_import('odl'): 0.39207029342651367 seconds
• test_import('numpy'): 0.0570526123046875 seconds
• test_import('matplotlib.pyplot'): 0.18544530868530273 seconds

Results seem to vary a bit depending on IPython vs. plain Python console and IDE vs. plain console. Times for importing ODL:

• Pure Python console in Spyder: 0.498138427734375 s
• IPython console in Spyder: 0.2175889015197754 s

Probably, caching also plays a role. Earlier today times were much worse. If I remove all .pyc files, ODL imports in 0.5289430618286133 seconds in a non-IDE IPython console.

[adler-j]

This is now significantly improved. Closing.