Support for custom priors via Prior class #488
Conversation
| @@ -245,10 +259,9 @@ def build_model(self, X, y, coords): | |||
| self.add_coords(coords) | |||
| X = pm.Data("X", X, dims=["obs_ind", "coeffs"]) | |||
| y = pm.Data("y", y, dims="obs_ind") | |||
| beta = pm.Normal("beta", 0, 50, dims="coeffs") | |||
| sigma = pm.HalfNormal("sigma", 1) | |||
There was a problem hiding this comment.
sigma will not be in the model anymore but rather, y_hat_sigma based on the default name generation. Is that breaking change an issue? There is a workaround for this if needed
There was a problem hiding this comment.
I don't think this is a big deal. I pushed a fix which make tests pass a9f821c (well looks like there is one failing doctest)
| @@ -286,9 +303,8 @@ def build_model(self, X, y, coords): | |||
| X = pm.Data("X", X, dims=["obs_ind", "coeffs"]) | |||
| y = pm.Data("y", y[:, 0], dims="obs_ind") | |||
| beta = pm.Dirichlet("beta", a=np.ones(n_predictors), dims="coeffs") | |||
| sigma = pm.HalfNormal("sigma", 1) | |||
There was a problem hiding this comment.
same breaking change concern
There was a problem hiding this comment.
This is very cool!
As far as I can tell, these changes are currently entirely invisible to the end user, right?
Before we do a new release I will follow up with another PR that adds some documentation. This will most likely be in the form of just giving a couple of worked examples
Main request is to add custom-ness to the Dirichlet for the WeightedSumFitter class. Like I say in a comment, I mostly see people wanting to customise the hyperparams, not the distribution itself.
| @@ -237,6 +248,11 @@ class LinearRegression(PyMCModel): | |||
| Inference data... | |||
| """ # noqa: W605 | |||
|
|
|||
| default_priors = { | |||
There was a problem hiding this comment.
need to add @property decorator here? Or is that remembered from it being done in the PyMCModel base class?
Getting an Pylance warning: Type "dict[str, Prior]" is not assignable to declared type "property"
There was a problem hiding this comment.
What line of code bring that on? Maybe having a setter will help?
| @@ -245,10 +259,9 @@ def build_model(self, X, y, coords): | |||
| self.add_coords(coords) | |||
| X = pm.Data("X", X, dims=["obs_ind", "coeffs"]) | |||
| y = pm.Data("y", y, dims="obs_ind") | |||
| beta = pm.Normal("beta", 0, 50, dims="coeffs") | |||
| sigma = pm.HalfNormal("sigma", 1) | |||
There was a problem hiding this comment.
I don't think this is a big deal. I pushed a fix which make tests pass a9f821c (well looks like there is one failing doctest)
| @@ -286,9 +303,8 @@ def build_model(self, X, y, coords): | |||
| X = pm.Data("X", X, dims=["obs_ind", "coeffs"]) | |||
| y = pm.Data("y", y[:, 0], dims="obs_ind") | |||
| beta = pm.Dirichlet("beta", a=np.ones(n_predictors), dims="coeffs") | |||
| sigma = pm.HalfNormal("sigma", 1) | |||
| @@ -286,9 +305,8 @@ def build_model(self, X, y, coords): | |||
| X = pm.Data("X", X, dims=["obs_ind", "coeffs"]) | |||
| y = pm.Data("y", y[:, 0], dims="obs_ind") | |||
| beta = pm.Dirichlet("beta", a=np.ones(n_predictors), dims="coeffs") | |||
There was a problem hiding this comment.
We definitely want a custom prior for the Dirichlet. I think the Dirichlet would be used always (or nearly always), but there are plenty of real world use cases where the user might want to change the hyper parameters (currently a=np.ones(n_predictors)).
There was a problem hiding this comment.
Alright. since it is function of data, we will have to handle differently
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #488 +/- ##
==========================================
+ Coverage 94.40% 94.42% +0.01%
==========================================
Files 29 29
Lines 2075 2081 +6
==========================================
+ Hits 1959 1965 +6
Misses 116 116 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
|
tests pass :) |
This using
pymc-extrasto allow for custom priors. Might need to adjust the remaining a bit to work.