Make the default shift size h of finite_diff and spsa_grad depend on shots
#5568
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
dwierichs
commented
Apr 24, 2024
Comment on lines
-361
to
+369
| >>> qml.gradients.finite_diff(circuit, h=10e-2)(params) | ||
| >>> qml.gradients.finite_diff(circuit, h=0.1)(params) |
dwierichs
commented
Apr 24, 2024
| @@ -787,6 +779,7 @@ def test_var_expectation_values(self, approx_order, strategy, validate): | |||
| def test_prob_expectation_values(self, approx_order, strategy, validate): | |||
| """Tests correct output shape and evaluation for a tape | |||
| with prob and expval outputs""" | |||
| np.random.seed(239) | |||
There was a problem hiding this comment.
Failed randomly while testing locally, so I add a seed.
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #5568 +/- ##
==========================================
- Coverage 99.69% 99.68% -0.01%
==========================================
Files 410 410
Lines 38247 37966 -281
==========================================
- Hits 38130 37848 -282
- Misses 117 118 +1 ☔ View full report in Codecov by Sentry. |
dwierichs
added
the
do not merge ⚠️
|
It was decided not to implement this change. Good to close @trbromley @co9olguy ? |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Context:
Finite differences use a shift parameter
hby which input parameters are shifted. The optimal choice of this parameter (bias-variance tradeoff) is not known a priori, but it is well-known that analytic and shot-based computations have very different optimal settings.The current default
1e-7is a poor choice for shot-based evaluation because of the large variance incurred by sampling. Forshots=Noneit is a reasonable setting.The same holds for
spsa_grad, which uses a finite difference recipe under the hood as well.Description of the Change:
Introduce helper functions to
finite_diffandspsa_gradthat sets the default value ofhto0.1when the differentiated tape hasshots!=None. Ifshots=None, the previous default values (1e-7forfinite_diff,1e-5forspsa_grad) are maintained, and of course a user-specifiedhoverwrites these default values.Benefits:
Reasonable results for the gradient transforms when using shots and not specifying
hmanually.In particular, when
param_shiftfalls back tofinite_diff, nohvalue can be specified.Possible Drawbacks:
Defaults are set arbitrary, but so they were before.
Related GitHub Issues:
#904
Regarding this issue: A way more involved relationship between shots and shift size is proposed in the discussion there.
Given that unknown parameters would be needed to implement the equation suggested there, and that it only holds for a particular finite difference recipe (and thus not for other
approx_order,norstrategysettings), it seems infeasible to implement such a complex heuristic. This PR is a coarse-grained fix preventing PennyLane from running almost-surely unreasonable circuits by default.@glassnotes @co9olguy
[sc-4095]