Move matplotlib imports to runtime #5485
Conversation
This commit moves all the matplotlib imports to runtime. Prior to this commit when importing the qiskit namespace matplotlib would always be imported if it was installed. This however adds a noticeable bottleneck to the import time, even if visualization isn't used. To avoid this overhead we can lazily import matplotlib at runtime when it is available. This commit makes this change so that matplotlib is only imported at runtime when visualization methods using it are called. This does slow down the first visualization function call using matplotlib, but since visualization isn't a performance critical path this an acceptable tradeoff. For backwards compibility the HAS_MATPLOTLIB module attribute has changed from a bool to an object of a new class HasMatplotlibWrapper, which implements a single method __bool__ which returns true if matplotlib is installed, otherwise it's false. However it does this at runtime instead of at import time. This way it be used as a replacement for the previous HAS_MATPLOTLIB boolean variable but without forcing matplotlib be imported with the entire package.
mtreinish
requested review from
maddy-tod,
nkanazawa1989,
nonhermitian and
a team
as code owners
|
Looking at the import times without this PR the profile looks like:
Then applying this PR speeds up the import noticeably to:
After this PR the only easy thing left is to remove the fastjsonschema usage. That is just a matter of time since it is deprecated (although only on master so the timer starts after the next release). |
| @@ -175,7 +173,7 @@ def pulse_drawer(data: Union[Waveform, Union[Schedule, Instruction]], | |||
|
|
|||
| if get_backend() in ['module://ipykernel.pylab.backend_inline', | |||
| 'nbAgg']: | |||
| _matplotlib.plt.close(image) | |||
| plt.close(image) | |||
There was a problem hiding this comment.
Is calling close on qiskit.visualization.pulse.matplotlib (which it was before) the same as calling it on matplotlib (which it is now)?
There was a problem hiding this comment.
It should be, matplotlib.pyplot is a global state instance.
That being said this comment reminded me that I need to check the pulse v2 drawer because I originally wrote this before that merged. Let me do a quick pass on that now and verify that matplotlib doesn't get imported at import time from there.
There was a problem hiding this comment.
ok, just double checked and it looks like @nkanazawa1989 already made the matplotlib imports occur at runtime for the v2 drawer so there aren't any changes I need to make there.
In Qiskit#5582 we added a script that is used to ensure we don't import slow optional dependencies in the default 'import qiskit' path. Since that PR merged we've managed to remove a few more packages from the default import path (see Qiskit#5619, Qiskit#5485, and Qiskit#5620) which we should add the same checks to to ensure we don't regress and accidently start importing them again in the default path. This commit adds all these packages to the list of imports not allowed as part of 'import qiskit'.
In #5582 we added a script that is used to ensure we don't import slow optional dependencies in the default 'import qiskit' path. Since that PR merged we've managed to remove a few more packages from the default import path (see #5619, #5485, and #5620) which we should add the same checks to to ensure we don't regress and accidently start importing them again in the default path. This commit adds all these packages to the list of imports not allowed as part of 'import qiskit'.
Summary
This commit moves all the matplotlib imports to runtime. Prior to this
commit when importing the qiskit namespace matplotlib would always be
imported if it was installed. This however adds a noticeable bottleneck
to the import time, even if visualization isn't used. To avoid this
overhead we can lazily import matplotlib at runtime when it is
available. This commit makes this change so that matplotlib is only
imported at runtime when visualization methods using it are called. This
does slow down the first visualization function call using matplotlib,
but since visualization isn't a performance critical path this an
acceptable tradeoff.
For backwards compibility the HAS_MATPLOTLIB module attribute has
changed from a bool to an object of a new class HasMatplotlibWrapper,
which implements a single method
__bool__which returns true ifmatplotlib is installed, otherwise it's false. However it does this at
runtime instead of at import time. This way it be used as a replacement
for the previous HAS_MATPLOTLIB boolean variable but without forcing
matplotlib be imported with the entire package.
Details and comments
Related to: #5100