Keep 0 dimensional parameters as 0 dimensional instead of 1 dimension…

…al (#575)

* start to reimplement predictions

* add playground notebooks, so we don't loose the work

* More advances

* Univariate families now use xarray

* Advances in multivariate families, there are bugs

* categorical family bug fixed

* First 'working' version of new predictions

* Fix shape in predictions with offset

* Final tweaks. Run all examples

* 1d caveats

* shape[1] instead of just shape

bambi/backend/pymc.py

@@ -158,6 +158,7 @@ def _build_common_terms(self, spec):
 
             # Column vector of coefficients and design matrix
             coefs = at.concatenate(coefs)
+
             data = np.hstack(columns)
 
             # If there's an intercept, center the data

bambi/backend/terms.py

@@ -48,7 +48,9 @@ def build(self, spec):
         if dims:
             coef = distribution(label, dims=dims, **args)
         else:
-            coef = distribution(label, shape=data.shape[1], **args)
+            shape = None if data.shape[1] == 1 else data.shape[1]
+            coef = distribution(label, shape=shape, **args)
+            coef = at.atleast_1d(coef)  # If only a single numeric column it wont be 1d
 
         # Prepends one dimension if response is multivariate categorical and predictor is 1d
         if response_dims and len(dims) == 1: