use numpy indexing in cdap.make_interactions

Once again we get significant speedups by working with NumPy arrays
instead of Pandas structures.

activitysim/cdap/cdap.py

@@ -48,24 +48,42 @@ def make_interactions(people, hh_id_col, p_type_col):
     three_fmt = '{}{}{}'.format
     two = []
     three = []
+    two_perm_cache = {}
+    three_combo_cache = {}
+
+    for hh_id, df in people.groupby(hh_id_col, sort=False):
+        hh_size = len(df)
 
-    for hh, df in people.groupby(hh_id_col, sort=False):
         # skip households with only one person
-        if len(df) == 1:
+        if hh_size == 1:
             continue
 
-        ptypes = df[p_type_col]
+        ptypes = df[p_type_col].values
+        hh_idx = df.index.values
 
-        for pA, pB in itertools.permutations(df.index, 2):
-            two.append((pA, two_fmt(*ptypes[[pA, pB]])))
+        if hh_size in two_perm_cache:
+            two_perms = two_perm_cache[hh_size]
+        else:
+            two_perms = list(itertools.permutations(np.arange(hh_size), 2))
+            two_perm_cache[hh_size] = two_perms
+
+        two.extend(
+            (hh_idx[pA], two_fmt(*ptypes[[pA, pB]])) for pA, pB in two_perms)
 
         # now skip households with two people
-        if len(df) == 2:
+        if hh_size == 2:
             continue
 
-        for idx in itertools.combinations(df.index, 3):
-            combo = three_fmt(*ptypes[list(idx)])
-            three.extend((p, combo) for p in idx)
+        if hh_size in three_combo_cache:
+            three_combos = three_combo_cache[hh_size]
+        else:
+            three_combos = list(itertools.combinations(np.arange(hh_size), 3))
+            three_combo_cache[hh_size] = three_combos
+
+        three.extend(
+            (hh_idx[p], three_fmt(*ptypes.take(idx)))
+            for idx in three_combos
+            for p in idx)
 
     if two:
         two_idx, two_val = zip(*two)