ForwardDiff with Interval partials - #769
Conversation
|
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #769 +/- ##
==========================================
- Coverage 75.48% 75.35% -0.14%
==========================================
Files 32 32
Lines 3100 3112 +12
==========================================
+ Hits 2340 2345 +5
- Misses 760 767 +7 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
|
I am not sure in what context |
|
For the rest of the extension, the strategy has been to reimplement the short-circuiting behaviors, removing the short circuit altogether (that's why there is a version of This was a conscious choice: we unfortunately do not have a function testing if "this value can be discarded for the purpose of addition" (which is how ForwardDiff uses Therefore, if it is a reasonable amount of work, I would keep de-short-circuiting ForwardDiff for now. Finding a more general solution would be great, but it probably requires a deep discussion with the ForwardDiff team. |
Add power methods with decidable thin-zero checks for interval-valued partials. Support real, interval, and exact exponents through four nested Dual layers. Assisted-by: Claude Fable 5 <noreply@anthropic.com>
|
Sorry for the delay. The new version follows @Kolaru's recommended design. |
Benchmark Results (Julia v1)Time benchmarks
Memory benchmarks
|
This change supports differentiation of functions with interval-valued partials, like
Previously, those failed due to a short-circuit in ForwardDiff, triggering
This short-circuits the short-circuit, by specializing
Base.iszeroon Interval-containingForwardDiff.DualandForwardDiff.Partials. No behavior change in IntervalArithmetic itself is needed.