Miscellaneous Bugfixes for 0.9.14 #640
Conversation
Fix a bug in the conversion from unitary model members to error generator ones present in the new SPAM conversion logic. All that was needed here was to change the space for the to_dense calls so they were consistent. Additionally adds a docstring to convert_members_inplace and set_default_gauge_group for the explicit model class. For set_default_gauge_group I have also changed the default behavior for the set_default_gauge_group function. Previously this would default to the TrivialGaugeGroup for all or most of the valid lindblad types. But in this case the unitary group is a better (though imperfect for edge cases) choice.
Partial patch for gauge optimization when using the statevec evotype that arises from difference in minimal reps for gauge group elements and operations.
There is an incompatibility between the statevec evotype and the representations used in gauge optimization. For now turn off gauge optimization on the GST tests which use the term forward simulator. Doing this most straightforwardly required adding a new option for the GST driver function.
Fix a bug in deserializing models saved before the recent changes to allow saving fogi stores.
Patch the diamond distance reportable function to account for a change to the return call signature for the diamond distance when returning the value of the SDP solution.
Some recent changes resulted in serialization using legacy json and msgpack serialization protocols breaking. This functionality has been deprecated for over two years, and so spending energy patching this doesn't make sense. Rather, this is as good a time as any to finally expire this deprecation.
Cleans up additional reference to the now expired legacy json and msgpack serialization protocols.
Add in casting logic to set_errorgen_coefficients when the input lindblad dictionary has label types which do not match the object being updated. Also add handling for edge case where the update dictionary is empty.
Change the default for run_long_sequence_gst's gauge optimization to use stdgaugeopt
Update the default gauge optimization settings of run_long_sequence_gst_base to match the non base counterpart.
Fix a small logic bug that meant certain dictionary values weren't be retained properly.
Minor patches to address some unbound variable name and related errors in color box plots, and some of type mismatches in the workspace tables.
Regenerate reference files for test_packages module to account for recent changes.
Update gauge optimization labels to account for recent default gauge optimization suite change. Turn off gauge optimization for model types which do not properly support it.
We're encountering CI issues with this field, and some googling indicates this is because there have been some changes to the conventions for this and not all versions of the build utilities (like setuptools) support these changes across all versions yet.
coreyostrove
requested review from
pcwysoc,
a team,
rileyjmurray and
kmrudin
as code owners
|
@coreyostrove I've triggered a CI build that runs the tests for beta and master, changed to target this branch: https://github.com/sandialabs/pyGSTi/actions/runs/17054858542 If those tests pass then I'll approve. Makes more sense to go this route than have me review the individual changes. |
| Smx = s.transform_matrix; Si = s.transform_matrix_inverse | ||
| Tinv = _np.linalg.inv(self.factorops[0].to_dense("minimal")) | ||
| trans_eLT = _np.dot(Si, _np.dot(self.to_dense("minimal"), Smx)) | ||
| Tinv = _np.linalg.inv(self.factorops[0].to_dense("HilbertSchmidt")) |
There was a problem hiding this comment.
Why are we limiting this hack for state_space.dim <= 16?
There was a problem hiding this comment.
Presumably just because it is expensive to perform the dense operations required with 3 or more qubits, but is faster than whatever other code path is used for 2Q and 3Q. This is incidentally not the biggest problem with this function at the moment, though. See #620 for more on that.
Revert changes from when I was debugging.
Something changed in some package which is producing np float objects which didn't have serialization logic in nameddict. Add a cast to standard python float.
Numpy 2.3 expired this deprecation for tostring.
|
The beta tests on the runners caught a couple extra issues that I didn't get locally. One was an expired deprecation of the |
|
Relaunched the beta tests with those extra patches (still need to fix the cython environment variable). |
See if we can figure out why it is that the workflow is triggering the ubuntu and macos branch on windows.
Awesome, thanks! I'm going to keep plugging away at getting the workflow corrected for the no cython build on this branch and will merge it in once those are running correctly. |
This reverts commit 41dde39.
This reverts commit e89ca79.
Realistically we probably don't need to run against all OS combinations to achieve the coverage goals of this particular build. Save on some resources.
…nto bugfixes-914
Previous attempt was still running partial workflows on windows and macos.
And remove a couple unused imports from fogistore.py
Strip newlines and carriage returns from output before doing error checking.
Lower the maximum depth of the fisher information computation with instruments in unit test. Also a miscellaneous workflow change.
Fix a bug related to a mismatch between current behavior of DataSet.gate_labels and target model labels.
This PR consists of miscellaneous bug fixes and patches in preparation of the 0.9.14 release which were being caught by tests within the
test_packagessubmodule (which are run against pushes to beta and master). Here is a brief summary:convert_members_inplaceandset_default_gauge_group_for_member_typemethods of theExplicitModelclass.set_default_gauge_group_for_member_typefor error generator parameterized models. Previously this was configured to set the default gauge group for such models to theTrivialGaugeGroup. Now we set this to be theUnitaryGaugeGroup, as one would more likely expect.pygsti.report.reportablesrelated to a recent change in the return signature for thediamonddistfunction inoptools.NicelySerializableinterface) and msgpack. We formally deprecated this functionality just over two years and had these marked for removal, so any users who might've been using this have been getting warnings for some time. Rather than invest any time fixing this, it makes more sense to finally expire these deprecations.run_long_sequence_gstandrun_long_sequence_gst_base. At the same time the default gauge optimization suite used for these functions has been switched to 'stdgaugeopt'. This brings these functions in line with their OOP-based counterpartGateSetTomographywhich has used this gauge optimization suite as the default for a while.set_errorgen_coefficientsmethod ofComposedOpto add handling for the cases where the update dictionary is empty, or has keys which do not match theElementaryErrorgenLabeltype of the modelmember.test_packageshave been regenerated to account for recent default setting changes.With these updates we have the tests on beta now passing (at least when run locally on my system, TBD on runners).