Features/exact #712
Features/exact #712
Conversation
| @@ -215,7 +215,6 @@ def factor_approximation(self, factor: Factor) -> FactorApproximation: | |||
| for v in factor_dist.all_variables | |||
| }).prod( | |||
| *factor_mean_field.values(), | |||
| default=factor_dist, | |||
There was a problem hiding this comment.
This causes incorrect initialisation of the cavity dist if a variable is used by only one factor.
There was a problem hiding this comment.
It's quite common in our usage for variables to only be used by a single factor. Many factors have 'dangling' variables which are not shared
There was a problem hiding this comment.
Do you mean you've fixed that issue or newly introduced it?
There was a problem hiding this comment.
I think I fixed it.
The way that you solved the dangling variable problem was causing variance shrinkage in the case of dangling variables.
There was a problem hiding this comment.
This way the dangling variable has no cavity distribution associated with it in the FactorApproximation which is fine for a lot of factor optimisations.
If you require access to the cavity distribution as a prior then you shouldn't have a dangling variable. (The prior should be present as a second factor connected to the variable).
There was a problem hiding this comment.
Actually yeah I guess every variable at least has a PriorFactor attached and so would be shared.
There was a problem hiding this comment.
I guess this may of partly also been responsible for the variance shrinkage in my tests, as each 1D Gaussian dataset had 2 parameters which were only tied to one factor?
There was a problem hiding this comment.
A test for this happening would be to run a factor optimisation multiple times in a row for the same factor, if the variance shrinks each step, then you're doing something like this.
Codecov Report
@@ Coverage Diff @@
## main #712 +/- ##
==========================================
+ Coverage 82.00% 82.23% +0.22%
==========================================
Files 180 180
Lines 13264 13305 +41
==========================================
+ Hits 10877 10941 +64
+ Misses 2387 2364 -23 see 9 files with indirect coverage changes 📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
There was a problem hiding this comment.
Looks like it only failed because of a drop in code coverage. Is there a test or two you could write? I think the paths stuff is probably not working properly but it would be a bit trickier to write a decent test for that as you would have to mock up the optimisers.
Also some docs would be nice
Otherwise looks good
| self.paths = paths | ||
| if self.paths: | ||
| for optimiser in self.factor_optimisers.values(): | ||
| optimiser.paths = self.paths |
There was a problem hiding this comment.
We probably want each factor_optimiser to have a different paths to ensure that they do not clash.
If multiple DynestyStatic had the same paths, for example, we might see them loading each other's optimisation state.
You could use the create_child method and pass the name of each factor.
There was a problem hiding this comment.
I have no idea what paths are supposed to do but it was breaking some of my code, so I put in code paths so that it became optional. In a later commit I've set the paths attribute to be optionally set if it doesn't exist.
There was a problem hiding this comment.
Actually had to revert the change, if you don't have this line test_autofit/graphical/info/test_output.py fails.
There was a problem hiding this comment.
It's basically an object that handles saving and loading data for optimisations. We used to have path manipulation stuff litered all over the optimiser classes so we put it all in one place. It can also conveniently be swapped out for a database client so data is persisted directly into a SQL database rather than being saved in a directory structure.
| @@ -89,6 +90,9 @@ def __init__( | |||
| self.log_norm = log_norm | |||
| self._plates = self.sorted_plates if plates is None else plates | |||
|
|
|||
| def copy(self) -> "MeanField": | |||
| return type(self)(self) | |||
| return VariableData({v: dist.mean for v, dist in self.items()}) | ||
| return self.attr("mean") | ||
|
|
||
| @property | ||
| def variance(self): | ||
| return VariableData({v: dist.variance for v, dist in self.items()}) | ||
| return self.attr("variance") | ||
|
|
||
| @property | ||
| def std(self): | ||
| return VariableData({v: dist.std for v, dist in self.items()}) | ||
| return self.attr("std") | ||
|
|
||
| @property | ||
| def scale(self): | ||
| return VariableData({v: dist.scale for v, dist in self.items()}) | ||
| return self.attr("scale") |
| cavity_dist = factor_approx.cavity_dist | ||
| cavity_dist = factor_approx.cavity_dist.copy() | ||
| for v in factor_approx.mean_field.keys() - cavity_dist: | ||
| cavity_dist[v] = factor_approx.mean_field[v].zeros_like() |
There was a problem hiding this comment.
For exact_fits of FactorApproximations we need the full cavity_dist, so in this case we need to create zeros_like versions of the missing cavity_dists - e.g. for the Normal Distribution the zeros_like dist has infinite variance, note that its logpdf will be - inf so we don't want to evaluate it in general.
There was a problem hiding this comment.
Interesting. Could you write something along those lines in the code docs?
|
I don't know why it failed. It works on my machine... |
Updating graphical api to allow better exact updates.
Added graphical/regression/test_exact.py to test exact EP updates.
In that case an exact linear regression and exact probit update is used to build a linear classification model.