Speed up cloned geography assignment

changelog.d/fast-geography-collisions.changed

@@ -0,0 +1 @@
+Speed up cloned household geography assignment for large local-area calibration builds.

policyengine_us_data/calibration/clone_and_assign.py

@@ -296,36 +296,35 @@ def _sample(size, mask_slice=None, fixed_slice=None):
     # Clone 0: unrestricted draw
     indices[:n_records] = _sample(n_records, extreme_mask, fixed_states)
 
-    assigned_cds = np.empty((n_clones, n_records), dtype=object)
-    assigned_cds[0] = cds[indices[:n_records]]
+    _, cd_codes = np.unique(cds, return_inverse=True)
+    cd_codes = cd_codes.astype(np.int32, copy=False)
+    record_positions = np.arange(n_records)
+    used_cd_by_record = np.zeros((n_records, cd_codes.max() + 1), dtype=bool)
+    used_cd_by_record[record_positions, cd_codes[indices[:n_records]]] = True
 
     for clone_idx in range(1, n_clones):
         start = clone_idx * n_records
         clone_indices = _sample(n_records, extreme_mask, fixed_states)
-        clone_cds = cds[clone_indices]
-
-        collisions = np.zeros(n_records, dtype=bool)
-        for prev in range(clone_idx):
-            collisions |= clone_cds == assigned_cds[prev]
+        clone_cd_codes = cd_codes[clone_indices]
+        collisions = used_cd_by_record[record_positions, clone_cd_codes]
 
         for _ in range(50):
-            n_bad = collisions.sum()
+            n_bad = int(collisions.sum())
             if n_bad == 0:
                 break
             bad_mask = collisions
             if extreme_mask is not None and agi_probs is not None:
-                replacement = _sample(n_records, extreme_mask, fixed_states)
-                clone_indices[bad_mask] = replacement[bad_mask]
+                fixed_bad = fixed_states[bad_mask] if fixed_states is not None else None
+                replacement = _sample(n_bad, extreme_mask[bad_mask], fixed_bad)
             else:
-                replacement = _sample(n_records, fixed_slice=fixed_states)
-                clone_indices[collisions] = replacement[collisions]
-            clone_cds = cds[clone_indices]
-            collisions = np.zeros(n_records, dtype=bool)
-            for prev in range(clone_idx):
-                collisions |= clone_cds == assigned_cds[prev]
+                fixed_bad = fixed_states[bad_mask] if fixed_states is not None else None
+                replacement = _sample(n_bad, fixed_slice=fixed_bad)
+            clone_indices[bad_mask] = replacement
+            clone_cd_codes = cd_codes[clone_indices]
+            collisions = used_cd_by_record[record_positions, clone_cd_codes]
 
         indices[start : start + n_records] = clone_indices
-        assigned_cds[clone_idx] = clone_cds
+        used_cd_by_record[record_positions, clone_cd_codes] = True
 
     assigned_blocks = blocks[indices]
     return GeographyAssignment(