[Capture] Add finite differences jvps #6853
Conversation
|
Hello. You may have forgotten to update the changelog!
|
…pennylane into finite-diff-capture
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #6853 +/- ##
=======================================
Coverage 99.54% 99.55%
=======================================
Files 477 477
Lines 45246 45273 +27
=======================================
+ Hits 45042 45070 +28
+ Misses 204 203 -1 ☔ View full report in Codecov by Sentry. |
| return qml.expval(qml.Z(0)) | ||
|
|
||
| with pytest.warns(UserWarning, match="Detected float32 parameter with finite differences."): | ||
| jax.grad(circuit)(jnp.array(0.5, dtype=jnp.float32)) |
There was a problem hiding this comment.
Should we be using qml.grad in the tests?
There was a problem hiding this comment.
Potential over-testing suggestion: or maybe parametrizing over qml and jax diff functions
There was a problem hiding this comment.
This is a jax specific test.
| jax.grad(circuit)(jnp.array(0.5, dtype=jnp.float32)) | ||
|
|
||
|
|
||
| class TestGradients: |
There was a problem hiding this comment.
Non blocking questions:
Do we have tests to validate that the correct diff method is actually being used? Does it matter if we do?
There was a problem hiding this comment.
I have a test at the bottom where I double check the captured jaxpr.
Co-authored-by: Mudit Pandey <[email protected]>
Co-authored-by: Andrija Paurevic <[email protected]>
Context:
Currently, program capture only supports differentiation with backprop. This is only supported on
default.qubitDescription of the Change:
Adds the capability to take finite difference derivatives.
Benefits:
We can now differentiate anything, and we can differentiate with the lightnings.
Possible Drawbacks:
finite diff tends to be noisy to the point of uselessness.
Related GitHub Issues:
[sc-82167]