Fix cyclic list handling in set_rnd (#9056)

Track container identities before recursion and preserve seed advancement across cyclic references.

RED→GREEN: RecursionError and 42 != 43 → 9 focused tests passing. Full min-dependency suite: 9655 passed, 2492 skipped.

Author: ousamabenyounes

monai/data/utils.py

@@ -700,10 +700,14 @@ def set_rnd(obj, seed: int, _seen: set[int] | None = None) -> int:
     if _seen is None:
         _seen = set()
     if isinstance(obj, (tuple, list)):  # ZipDataset.data is a list
-        _seed = seed
+        if id(obj) in _seen:
+            return seed
+        _seen.add(id(obj))
+        has_randomizable = False
         for item in obj:
-            _seed = set_rnd(item, seed=seed, _seen=_seen)
-        return seed if _seed == seed else seed + 1  # return a different seed if there are randomizable items
+            item_seed = set_rnd(item, seed=seed, _seen=_seen)
+            has_randomizable = has_randomizable or item_seed != seed
+        return seed + 1 if has_randomizable else seed
     if not hasattr(obj, "__dict__"):
         return seed  # no attribute
     if id(obj) in _seen:

tests/data/test_dataloader.py

@@ -20,6 +20,7 @@
 from parameterized import parameterized
 
 from monai.data import CacheDataset, DataLoader, Dataset, ZipDataset
+from monai.data.utils import set_rnd
 from monai.transforms import Compose, DataStatsd, Randomizable, SimulateDelayd
 from monai.utils import convert_to_numpy, set_determinism
 from tests.test_utils import assert_allclose
@@ -28,6 +29,11 @@
 
 TEST_CASE_2 = [[{"label": torch.as_tensor([[3], [2]])}, {"label": np.asarray([[1], [2]])}]]
 
+_CYCLIC_DATASET_SIZE = 4
+_CYCLIC_BATCH_SIZE = 1
+_CYCLIC_NUM_WORKERS = 0
+_CYCLIC_TEST_SEED = 42
+
 
 class TestDataLoader(unittest.TestCase):
     def test_values(self):
@@ -117,18 +123,50 @@ def __init__(self):
         self.cfg = parent
 
     def __len__(self):
-        return 4
+        return _CYCLIC_DATASET_SIZE
 
     def __getitem__(self, index):
         return torch.tensor([index])
 
 
+class _SeedRecorder:
+    def __init__(self):
+        self.seed = None
+
+    def set_random_state(self, seed):
+        self.seed = seed
+
+
 class TestLoaderRecursion(unittest.TestCase):
     def test_cyclic_reference_no_recursion(self):
         # Constructing the loader seeds the dataset (num_workers=0). A reference cycle in the
         # dataset's attributes must not raise RecursionError while walking the object graph.
-        dataloader = DataLoader(_CyclicConfigDataset(), batch_size=1, num_workers=0, shuffle=False)
-        self.assertEqual(len(list(dataloader)), 4)
+        dataloader = DataLoader(
+            _CyclicConfigDataset(), batch_size=_CYCLIC_BATCH_SIZE, num_workers=_CYCLIC_NUM_WORKERS, shuffle=False
+        )
+        self.assertEqual(len(list(dataloader)), _CYCLIC_DATASET_SIZE)
+
+    def test_cyclic_list_reference_no_recursion(self):
+        """Test seeding a dataset whose configuration list contains itself."""
+        dataset = _CyclicConfigDataset()
+        dataset.cfg = []
+        dataset.cfg.append(dataset.cfg)
+        dataloader = DataLoader(dataset, batch_size=_CYCLIC_BATCH_SIZE, num_workers=_CYCLIC_NUM_WORKERS, shuffle=False)
+        self.assertEqual(len(list(dataloader)), _CYCLIC_DATASET_SIZE)
+
+    def test_cyclic_list_preserves_seed_advancement(self):
+        """Test a cyclic list does not erase seed advancement from an earlier item."""
+        dataset = _CyclicConfigDataset()
+        nested_randomizable = _SeedRecorder()
+        following_randomizable = _SeedRecorder()
+        dataset.cfg = [nested_randomizable]
+        dataset.cfg.append(dataset.cfg)
+        dataset.following_randomizable = following_randomizable
+
+        set_rnd(dataset, seed=_CYCLIC_TEST_SEED)
+
+        self.assertEqual(nested_randomizable.seed, _CYCLIC_TEST_SEED)
+        self.assertEqual(following_randomizable.seed, _CYCLIC_TEST_SEED + 1)
 
 
 if __name__ == "__main__":