WIP: TVD and other objectives in GST #649

rileyjmurray · 2025-09-10T17:53:18Z

Notes to self (updated sporadically):

There should be a way of adding custom objectives to pygsti without changing its source code. I should have a unit test that does this.
At some point I needed to make changes to copalayout.py to resolve an error in running an example notebook. The changes and notebook were first introduced in the commit with this comment. I've since reverted the changes to copalayout.py because it seems they aren't needed anymore.

…me a file in test/performance so that it doesnt get discovered by testing frameworks like pytest

…t commit will simpify further, which may have consequences from the perspective of floating point arithmetic. Checking in NOW in case we want to revert to this version for numerical reasons.

…eparately, modify docstring for run_long_sequence_gst to be consistent with its implementation (including leaving notes for later).

…uge_group. Having the casting occur here makes sense, since a model has access to the associated StateSpace, Basis, and Evotype.

…tchboard.objfn_builder was used when switchboard.objfn_builder_modvi should have been used

…g in develop)

…nse of being possibly-incorrect

…ectiveFunction with an implementation that works for five of its eight derived classes. Removed the separate implementations for those five derived classes. Introduce a templated docstring for TimeIndependentMDCObjectiveFunction and a decorator to reduce duplication among the nine total __init__ functions of this class and its eight derived classes .

…uilders has no derived classes.

… setting a module-level variable

…option of preserving parameterization. I set the default to False, since this seems to be needed for existing germ selection tests.

… if the model is serialized then it can be deserialized correctly.

…e a collision that puts correctness at risk

…be near zero for correctness. Tests now pass even when optools.py.__SCALAR_TOL_EXPONENT__ = -1, which is reasonable to set when trying to suppress warnings or certain errors

…when (de)serializing a model with gate labels like `GCNOT` and `GIX`"

…nges needed for test_gaugeopt_correctness.

nkoskelo · 2025-09-11T15:25:54Z

pygsti/data/datasetconstruction.py

+        interp_weights[interp_weights < p] = 0.0
+        interp_weights[interp_weights > 0] = 1.0 
+    else:
+        interp_weights = p * _np.ones(num_circuits)


Are we allowing p to be a arraylike here, so that one can fully specify the distribution of weights?

nkoskelo · 2025-09-11T15:34:20Z

pygsti/protocols/gst.py

+        objective : {'logl', 'chi2', 'tvd'}, optional
            Whether to create builders for maximum-likelihood or minimum-chi-squared GST.


Comment needs to be updated to include total variational distance GST.

nkoskelo · 2025-09-11T15:38:22Z

pygsti/protocols/gst.py

+            import warnings
+            warnings.warn(f'Trying to create an objective function from non-string specification, "{objective}". \
+                           \nSupport for this kind of specification is experimental!')
+            iteration_builders = [chi2_builder]


Do not need line 796.

nkoskelo · 2025-09-11T15:43:05Z

pygsti/report/section/summary.py

-            switchboard.modvi_ds, switchboard.mdl_final_modvi,
+            switchboard.objfn_builder_modvi,
+            switchboard.circuits_final,
+            switchboard.modvi_ds, switchboard.mdl_current_modvi,


switchboard.mdl_current_modvi seems inaccurate. This function is for the final_model_fit. Thus, shouldn't the appropriate model be switchboard.mdl_final_modvi?

nkoskelo · 2025-09-11T15:49:14Z

pygsti/tools/optools.py

@@ -99,6 +110,18 @@ def assert_hermitian(mat):
    assert_hermitian(a)
    assert_hermitian(b)

+    def check_unit_trace(mat):


Do we want check_unit_trace to be only callable within _flat_mut_blks?

nkoskelo · 2025-09-11T16:06:09Z

pygsti/drivers/longsequence.py

+        that we never reference in the code (truncScheme, estimate_label, and
+        missingDataAction).
+
+        - objective = typically, a string in {'chi2', 'logl', 'tvd'}. But this can


Add the Literal typing to the parameter objective.

These arguments are only passed in the advanced_options dict. I can add the type annotation that advanced_options is dict[str,Any]. I'll leave other type annotations of this function to later work.

nkoskelo · 2025-09-11T16:54:40Z

pygsti/objectivefns/objectivefns.py

-    def create_from(cls, objective='logl', freq_weighted_chi2=False):
+    # This was a classmethod, but I made it static because this class has no derived classes.
+    @staticmethod
+    def create_from(objective='logl', freq_weighted_chi2=False, callable_penalty=None):


objective should be typed as a Literal.

nkoskelo

Some items for consideration.

rileyjmurray · 2025-09-11T17:50:12Z

pygsti/layouts/copalayout.py

I needed to make these changes to resolve exceptions that were getting raised. I need to find an example showing how to trigger the error without these changes.

…of a TypeError from CircuitOutcomeProbabilityArrayLayout.indices_and_outcomes_for_index)

…ityArrayLayout.__iter__ (needed to convert slice to iterable)

rileyjmurray and others added 30 commits December 16, 2024 12:56

initial commit simplifying TimeIndependentMDCObjectiveFunction

bc64ecb

bugfixes

b6e540f

improve readability

8f59891

inline comment

986e528

add pytest dispatcher for MPI test in test_packages

ccf17ef

move simple MPI test from test/test_packages into test/unit, and rena…

5959a53

…me a file in test/performance so that it doesnt get discovered by testing frameworks like pytest

implement changes discussed in 12/17 dev meeting

19316be

expand scope of exception handling

9b73727

add separate TimeIndpendentMDCObjectiveFunction.dlsvec back

fb4818c

resolve an infuriating sign error. Will probably want to clean this up.

122fd11

check in slightly simpler (but still complicated) implementation. Nex…

2ef13e2

…t commit will simpify further, which may have consequences from the perspective of floating point arithmetic. Checking in NOW in case we want to revert to this version for numerical reasons.

simplifications complete

1c23251

add comment

6e9d366

initial changes to RawTVDFunction

7d20e53

revert unncessary change

d15110e

fix a silly bug and implement per-gate weighting for TVD

b0e72b5

Merge remote-tracking branch 'origin' into TVD2

2719da0

Merge branch 'develop' into TVD

c33304a

add option handling for TVD objective in user-facing GST functions. S…

f8fea8e

…eparately, modify docstring for run_long_sequence_gst to be consistent with its implementation (including leaving notes for later).

add string-casting to the setter method of ExplicitOpModel.default_ga…

a67bece

…uge_group. Having the casting occur here makes sense, since a model has access to the associated StateSpace, Basis, and Evotype.

add the option of skipping report sections in construct_standard_report

2ed1c21

fix LIKELY bug in SummarySection.final_model_fit_histogram, where swi…

d028a77

…tchboard.objfn_builder was used when switchboard.objfn_builder_modvi should have been used

bugfix

48fb7ae

check in chnages for example (and getting example to run after mergin…

14c1a51

…g in develop)

make entry point for final-iteration objective functions more flexible

5de61b8

remove commented-out code that was trying to be efficient at the expe…

99619f1

…nse of being possibly-incorrect

fix silly bugs

4d09819

correct log message in simplerlm.py. Add a comment about a TODO.

8d3bdef

make classmethods of GSTObjFnBuilders static methods, since GSTObjFnB…

ac67824

…uilders has no derived classes.

rileyjmurray added 16 commits September 3, 2025 10:24

make model.copy() more robust

4dc24d0

make ExplicitOpModel.randomize_with_unitary parameterization-preserving.

4cf4a3e

changes to ComposedPOVM and ComposedState so the previous commit works

ca892e5

propagate change to DenseState._from_memoized_dict from master branch

12d1ffb

smarter unittest decorators

5876e95

make it possible to control __SCALAR_TOL__ in the fidelty function by…

c268571

… setting a module-level variable

try-catch for more robust model copying

e12280d

guard to prevent silent errors during deserialization

87f91de

revise ExplicitOpModel.randomize_with_unitary so that users have the …

f6b1683

…option of preserving parameterization. I set the default to False, since this seems to be needed for existing germ selection tests.

add a guard in _create_explicit_model_from_expressions to ensure that…

3c2580e

… if the model is serialized then it can be deserialized correctly.

change error checking so we only raise an error if we actually observ…

d1a9f58

…e a collision that puts correctness at risk

introduce private tolerance parameter in fidelity(...) that needs to …

ac89712

…be near zero for correctness. Tests now pass even when optools.py.__SCALAR_TOL_EXPONENT__ = -1, which is reasonable to set when trying to suppress warnings or certain errors

bugfix in circuit parsing from string. Discovered the underlying bug …

db5e406

…when (de)serializing a model with gate labels like `GCNOT` and `GIX`"

add some circuitlist generation utilities for model locking

509b94f

complete merge

8723928

less invasive change to implementation of ExplicitOpModel.copy(). Cha…

4cdcf78

…nges needed for test_gaugeopt_correctness.

rileyjmurray mentioned this pull request Sep 10, 2025

WIP: TVD objective in GST #506

Closed

nkoskelo reviewed Sep 11, 2025

View reviewed changes

rileyjmurray commented Sep 11, 2025

View reviewed changes

rileyjmurray added 4 commits September 11, 2025 11:07

type annotation and comment fix

5e08e32

revert seemingly unnecessary change to copalayout.py (in the context …

9c64c9d

…of a TypeError from CircuitOutcomeProbabilityArrayLayout.indices_and_outcomes_for_index)

remove unused variable

09d7046

add type annotation and fix now-obvious bug in CircuitOutcomeProbabil…

cc687dd

…ityArrayLayout.__iter__ (needed to convert slice to iterable)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

WIP: TVD and other objectives in GST #649

WIP: TVD and other objectives in GST #649

rileyjmurray commented Sep 10, 2025 •

edited

Loading

Uh oh!

nkoskelo Sep 11, 2025

Uh oh!

nkoskelo Sep 11, 2025

Uh oh!

nkoskelo Sep 11, 2025

Uh oh!

nkoskelo Sep 11, 2025

Uh oh!

nkoskelo Sep 11, 2025

Uh oh!

nkoskelo Sep 11, 2025

Uh oh!

rileyjmurray Sep 11, 2025

Uh oh!

nkoskelo Sep 11, 2025

Uh oh!

nkoskelo left a comment

Uh oh!

rileyjmurray Sep 11, 2025

Uh oh!

Uh oh!

		objective : {'logl', 'chi2', 'tvd'}, optional
		Whether to create builders for maximum-likelihood or minimum-chi-squared GST.

WIP: TVD and other objectives in GST #649

Are you sure you want to change the base?

WIP: TVD and other objectives in GST #649

Conversation

rileyjmurray commented Sep 10, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nkoskelo left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

rileyjmurray commented Sep 10, 2025 •

edited

Loading