Correct handling of linear model fitting failure (labsyspharm#187)

Now we look at the determinant of the matrix of coefficients, not just whether all the individual values are zero. Also the intercept is now zeroed out which fixes the single tile placed far from the rest (even though this was a sign of complete EdgeAligner failure). A warning is emitted in the failure case to alert the user of the problem.
openUC2 · Apr 7, 2023 · 0176ba5 · 0176ba5
1 parent 6eeda68
commit 0176ba5
Showing 1 changed file with 12 additions and 3 deletions.
diff --git a/ashlar/reg.py b/ashlar/reg.py
@@ -532,7 +532,7 @@ def check_overlaps(self):
             warn_data("Some neighboring tiles have zero overlap.")
 
     def compute_threshold(self):
-        if self.max_error:
+        if self.max_error is not None:
             if self.verbose:
                 print("    using explicit error threshold")
             return
@@ -674,9 +674,18 @@ def fit_model(self):
         cc0 = list(components[0])
         self.lr = sklearn.linear_model.LinearRegression()
         self.lr.fit(self.metadata.positions[cc0], self.positions[cc0])
-        # Fix up degenerate transform matrix (e.g. when we have only one tile).
-        if (self.lr.coef_ == 0).all():
+        # Fix up degenerate transform matrix. This happens when the spanning
+        # tree is completely edgeless or cc0's metadata positions fall in a
+        # straight line. In this case we fall back to the identity transform.
+        if np.linalg.det(self.lr.coef_) < 1e-3:
+            # FIXME We should probably exit here, not just warn. We may provide
+            # an option to force it anyway.
+            warn_data(
+                "Could not align enough edges, proceeding anyway with original"
+                " stage positions."
+            )
             self.lr.coef_ = np.diag(np.ones(2))
+            self.lr.intercept_ = np.zeros(2)
         # Adjust position of remaining components so their centroids match
         # the predictions of the model.
         for cc in components[1:]: