Fix RecursionError in set_rnd on cyclic object graphs (#8087)

monai/data/utils.py

@@ -686,28 +686,36 @@ def worker_init_fn(worker_id: int) -> None:
     set_rnd(worker_info.dataset, seed=worker_info.seed)  # type: ignore[union-attr]
 
 
-def set_rnd(obj, seed: int) -> int:
+def set_rnd(obj, seed: int, _seen: set[int] | None = None) -> int:
     """
     Set seed or random state for all randomizable properties of obj.
 
     Args:
         obj: object to set seed or random state for.
         seed: set the random state with an integer seed.
+        _seen: internal set of already-visited object ids, used to guard against
+            infinite recursion on cyclic object graphs (e.g. OmegaConf/Hydra
+            configs whose child nodes back-reference their parent, see issue #8087).
     """
+    if _seen is None:
+        _seen = set()
     if isinstance(obj, (tuple, list)):  # ZipDataset.data is a list
         _seed = seed
         for item in obj:
-            _seed = set_rnd(item, seed=seed)
+            _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
     if not hasattr(obj, "__dict__"):
         return seed  # no attribute
+    if id(obj) in _seen:
+        return seed  # already visited: avoid infinite recursion on cyclic references
+    _seen.add(id(obj))
     if hasattr(obj, "set_random_state"):
         obj.set_random_state(seed=seed % MAX_SEED)
         return seed + 1  # a different seed for the next component
     for key in obj.__dict__:
         if key.startswith("__"):  # skip the private methods
             continue
-        seed = set_rnd(obj.__dict__[key], seed=seed)
+        seed = set_rnd(obj.__dict__[key], seed=seed, _seen=_seen)
     return seed
 
 

tests/data/test_dataloader.py

@@ -12,6 +12,7 @@
 from __future__ import annotations
 
 import sys
+import types
 import unittest
 
 import numpy as np
@@ -99,5 +100,36 @@ def test_zipdataset(self):
         assert_allclose(np.stack(output).flatten()[:7], np.array([594, 170, 594, 170, 594, 170, 524]))
 
 
+class _CyclicConfigDataset(torch.utils.data.Dataset):
+    """
+    Dataset holding an attribute whose object graph contains a reference cycle.
+
+    This mirrors OmegaConf/Hydra configs, whose child nodes hold a back-reference
+    to their parent node. Seeding such a dataset used to recurse forever in
+    ``monai.data.utils.set_rnd`` (see issue #8087).
+    """
+
+    def __init__(self):
+        parent = types.SimpleNamespace()
+        child = types.SimpleNamespace()
+        parent.child = child
+        child.parent = parent  # reference cycle, as in an OmegaConf parent/child graph
+        self.cfg = parent
+
+    def __len__(self):
+        return 4
+
+    def __getitem__(self, index):
+        return torch.tensor([index])
+
+
+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)
+
+
 if __name__ == "__main__":
     unittest.main()