[DataLoader2] Saving and restoring initial seed generator

test/dataloader2/test_mprs.py

@@ -197,6 +197,65 @@ def test_reading_service_limit(self, dp, n_workers, worker_prefetch_cnt, main_pr
  res.append(x)
  self.assertEqual(9, len(res))
 
+ def test_initial_epoch_checkpointing(self):
+ dp = IterableWrapper(range(20)).shuffle().sharding_filter()
+ rs = MultiProcessingReadingService(num_workers=2)
+
+ # Functional Test: Saving state before iterator is created
+ dl: DataLoader2 = DataLoader2(datapipe=dp, reading_service=rs)
+ dl.seed(1)
+ initial_state = dl.state_dict()
+ it1 = iter(dl)
+
+ restored_dl: DataLoader2 = DataLoader2.from_state(initial_state, rs) # type: ignore[arg-type]
+ self.assertEqual(list(it1), list(restored_dl))
+
+ dl.shutdown()
+ restored_dl.shutdown()
+
+ # Functional Test: Saving state after iterator is created
+ dl = DataLoader2(datapipe=dp, reading_service=rs)
+ dl.seed(1)
+ it1 = iter(dl)
+ initial_state = dl.state_dict()
+
+ restored_dl = DataLoader2.from_state(initial_state, rs) # type: ignore[arg-type]
+ self.assertEqual(list(it1), list(restored_dl))
+
+ dl.shutdown()
+ restored_dl.shutdown()
+
+ # Functional Test: Saving state after iterator is created and began iterating
+ dl = DataLoader2(datapipe=dp, reading_service=rs)
+ dl.seed(1)
+ it1 = iter(dl)
+ next(it1) # Starts iterating
+ initial_state = dl.state_dict()
+
+ restored_dl = DataLoader2.from_state(initial_state, rs) # type: ignore[arg-type]
+ restored_dl._initial_seed_generator = None
+
+ print(dl._seed_generator)
+ print(restored_dl._seed_generator)
+
+ print(dl._seed_generator == restored_dl._seed_generator)
+
+ restored_dl2 = DataLoader2.from_state(initial_state, rs) # type: ignore[arg-type]
+ restored_dl2._restore_checkpoint_beginning_of_epoch()
+
+ res1 = list(it1)
+ res2 = list(restored_dl)
+ res3 = list(restored_dl2)
+
+ print(res1)
+ print(res2)
+ print(res3)
+ # self.assertEqual(list(it1), list(restored_dl)[1:]) # Note skipping over 1st element from actual result
+
+ dl.shutdown()
+ restored_dl.shutdown()
+ restored_dl2.shutdown()
+
  # TODO: Test cases when there is official support of `pause` and `resume` with round-robin sharding
  # Currently, using sharding_round_robin raises a warning
  # def test_round_robin_dispatching_pause_limit(self):

torchdata/dataloader2/dataloader2.py

@@ -3,8 +3,7 @@
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
-
-
+import pickle
 import warnings
 
 from dataclasses import dataclass
@@ -20,6 +19,7 @@
 T_co = TypeVar("T_co", covariant=True)
 SERIALIZED_DATAPIPE_KEY_NAME = "serialized_datapipe"
 READING_SERVICE_STATE_KEY_NAME = "reading_service_state"
+RANDOMNESS_STATE_KEY_NAME = "randomness_state"
 
 
 @dataclass
@@ -185,6 +185,8 @@ def __init__(
  self._seed_generator: SeedGenerator = SeedGenerator()
  self._seed: Optional[int] = None
  self._reset_seed: bool = True
+ # Seed generator as of beginning of each epoch
+ self._initial_seed_generator: SeedGenerator = clone(self._seed_generator)
 
  def __iter__(self) -> DataLoader2Iterator[T_co]:
  r"""
@@ -207,6 +209,9 @@ def __iter__(self) -> DataLoader2Iterator[T_co]:
  else:
  self._seed_generator.seed()
 
+ # Saving initial seed generator state
+ self._initial_seed_generator = clone(self._seed_generator)
+
  if not self._adapted and self.reading_service is not None:
  if self.reading_service_state is None:
  self.datapipe = self.reading_service.initialize(self.datapipe)
@@ -278,10 +283,17 @@ def state_dict(self) -> Dict[str, Any]:
 
  # Serialize datapipe after applying adapters and before reading service adaption
  serialized_datapipe = serialize_datapipe(self._datapipe_before_reading_service_adapt)
+ serialized_randomness_state = (
+ self._seed,
+ self._reset_seed,
+ pickle.dumps(self._seed_generator),
+ pickle.dumps(self._initial_seed_generator),
+ )
 
  return {
  SERIALIZED_DATAPIPE_KEY_NAME: serialized_datapipe,
  READING_SERVICE_STATE_KEY_NAME: reading_service_state,
+ RANDOMNESS_STATE_KEY_NAME: serialized_randomness_state,
  }
 
  @classmethod
@@ -303,6 +315,12 @@ def from_state(
  reading_service=reading_service,
  )
  data_loader.reading_service_state = reading_service_state
+
+ randomness_state = state[RANDOMNESS_STATE_KEY_NAME]
+ data_loader._seed, data_loader._reset_seed = randomness_state[0], randomness_state[1]
+ data_loader._seed_generator = pickle.loads(randomness_state[2])
+ data_loader._initial_seed_generator = pickle.loads(randomness_state[3])
+
  return data_loader
 
  def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
@@ -329,12 +347,20 @@ def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
  self.datapipe = deserialized_datapipe
  self.reading_service_state = reading_service_state
 
+ randomness_state = state_dict[RANDOMNESS_STATE_KEY_NAME]
+ self._seed, self._reset_seed = randomness_state[0], randomness_state[1]
+ self._seed_generator = pickle.loads(randomness_state[2])
+ self._initial_seed_generator = pickle.loads(randomness_state[3])
+
  # re-initialize datapipe_adapter_fn and _datapipe_before_reading_service_adapt
  if self.datapipe_adapter_fns is not None:
  for adapter_fn in self.datapipe_adapter_fns:
  self.datapipe = adapter_fn(self.datapipe)
  self._datapipe_before_reading_service_adapt = clone(self.datapipe)
 
+ def _restore_checkpoint_beginning_of_epoch(self) -> None:
+ self._seed_generator = self._initial_seed_generator
+
  def _pause(self):
  if hasattr(self.reading_service, "_pause"):
  self._is_paused = True