Structured smoothing #69
Conversation
Codecov Report
@@ Coverage Diff @@
## master #69 +/- ##
==========================================
+ Coverage 82.04% 84.06% +2.02%
==========================================
Files 29 29
Lines 2389 2454 +65
==========================================
+ Hits 1960 2063 +103
+ Misses 429 391 -38
Continue to review full report at Codecov.
|
| slc = smooth(sdd) | ||
| plc = propagate_constants(sdd, remove_unary=true) | ||
| structplc = compile(StructLogicCircuit, vtr, plc) | ||
| sstructplc = smooth(structplc) |
There was a problem hiding this comment.
Why do we need to smooth it twice? Is it because propagage constants need smooth circuit?
There was a problem hiding this comment.
The smoothed logic circuit is totally orthogonal, it's just for testing. I'm using probabilistic equivalence checking to make sure the smoothing didn't somehow change the meaning of the circuit (that you get the same thing regardless of how you smooth).
There was a problem hiding this comment.
oh, makes sense, I see now, initially read line 22 wrong.
This adds smoothing for structured logic circuits, respecting the current vtree. Appears to be working - looking for more example tests that don't have dependencies in Probabilistic circuits.
I couldn't find a way around doing work at both AND and OR nodes, it's possible there is some unnecessary work going on.
As of right now this will only work on logic circuits due to typing issues. A future pr should add a "structured decomposable" trait which would allow this to work with structured prob nodes as well.