Fixes

src/dials/algorithms/correlation/analysis.py

@@ -67,7 +67,7 @@
     source=iotbx.phil.parse(
         """\
 cc_weights=sigma
-weights=standard_error
+weights=count
 """
     )
 )
@@ -150,8 +150,8 @@ def __init__(
         # If dimensions are optimised for clustering, need cc_weights=sigma
         # Otherwise results end up being nonsensical even for high-quality data
         # Outlier rejection was also found to be beneficial for optimising clustering dimensionality
-        if self.params.dimensions is Auto and self.params.cc_weights != "sigma":
-            raise ValueError("To optimise dimensions, cc_weights=sigma is required.")
+        # if self.params.dimensions is Auto and self.params.cc_weights != "sigma":
+        #    raise ValueError("To optimise dimensions, cc_weights=sigma is required.")
 
         self.cosym_analysis = CosymAnalysis(self.datasets, self.params)
 

src/dials/algorithms/symmetry/cosym/__init__.py

@@ -87,7 +87,7 @@
   .type = bool
   .short_caption = "Use curvatures"
 
-weights = count standard_error
+weights = *count standard_error
   .type = choice
   .short_caption = "Weights"
   .help = "If not None, a weights matrix is used in the cosym procedure."

src/dials/algorithms/symmetry/cosym/engine.py

@@ -99,9 +99,9 @@ def minimize_scitbx_lbfgs(
         max_iterations=max_iterations,
         max_calls=max_calls,
         traditional_convergence_test=True,
-        traditional_convergence_test_eps=1,
+        traditional_convergence_test_eps=0.001,
         drop_convergence_test_n_test_points=5,
-        drop_convergence_test_max_drop_eps=1.0e-5,
+        drop_convergence_test_max_drop_eps=1e-7,
         drop_convergence_test_iteration_coefficient=2,
     )
     result = lbfgs_with_curvs(

src/dials/algorithms/symmetry/cosym/target.py

@@ -331,7 +331,7 @@ def _compute_rij_wij_ccweights(self):
                         wij_matrix += wij
 
         rij_matrix = rij_matrix.toarray().astype(np.float64)
-        if wij_matrix is not None:
+        if self._weights:
             wij_matrix = wij_matrix.toarray().astype(np.float64)
             if self._weights == "standard_error":
                 # N.B. using effective n due to sigma weighting, which can be below 2
@@ -341,6 +341,11 @@ def _compute_rij_wij_ccweights(self):
                 se = np.sqrt((1 - np.square(rij_matrix[sel])) / (wij_matrix[sel] - 1))
                 wij_matrix = np.zeros_like(rij_matrix)
                 wij_matrix[sel] = 1 / se
+        else:
+            ## make wij matrix with ones on off-diagonal elements.
+            wij_matrix = wij_matrix.toarray().astype(np.float64)
+            sel = np.where(wij_matrix > 0)
+            wij_matrix[sel] = 1
 
         return rij_matrix, wij_matrix
 
@@ -462,7 +467,25 @@ def sample_size(x, y):
                         + "\nIncreasing min_reflections may overcome this problem."
                     )
         else:
-            wij = None
+            wij = np.zeros_like(rij)
+            right_up = np.triu_indices_from(wij, k=1)
+
+            # For each correlation coefficient, set the weight equal to the size of
+            # the sample used to calculate that coefficient
+            pairwise_combos = itertools.combinations(np.isfinite(all_intensities), 2)
+
+            def sample_size(x, y):
+                pairs = np.count_nonzero(x & y)
+                if pairs < self._min_pairs:
+                    return 0
+                else:
+                    return pairs
+
+            wij[right_up] = list(itertools.starmap(sample_size, pairwise_combos))
+            sel = np.where(wij > 0)
+            wij[sel] = 1
+            # Symmetrise the wij matrix
+            wij += wij.T
 
         return rij, wij
 

tests/command_line/test_cosym.py

@@ -20,7 +20,7 @@
 @pytest.mark.parametrize(
     "space_group,engine,weights,cc_weights",
     [
-        (None, "scitbx", None, None),
+        (None, "scitbx", "count", "sigma"),
         ("P 1", "scipy", None, None),
         ("P 4", "scipy", "standard_error", "sigma"),
     ],