Call set_epoch for distributed batch samplers (#13396)

Co-authored-by: Jirka <[email protected]>
Co-authored-by: Rohit Gupta <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -243,15 +243,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 ### Fixed
 
 - Fixed `estimated_stepping_batches` requiring distributed comms in `configure_optimizers` for the `DeepSpeedStrategy` ([#13350](https://github.com/PyTorchLightning/pytorch-lightning/pull/13350))
-
-
 - Fixed bug with Python version check that prevented use with development versions of Python ([#13420](https://github.com/PyTorchLightning/pytorch-lightning/pull/13420))
-
-
 - Fixed Model Summary when using DeepSpeed Stage 3 ([#13427](https://github.com/PyTorchLightning/pytorch-lightning/pull/13427))
-
-
 - Fixed `pytorch_lightning.utilities.distributed.gather_all_tensors` to handle tensors of different dimensions ([#12630](https://github.com/PyTorchLightning/pytorch-lightning/pull/12630))
+- The loops now call `.set_epoch()` also on batch samplers if the dataloader has one wrapped in a distributed sampler ([#13396](https://github.com/PyTorchLightning/pytorch-lightning/pull/13396))
 
 
 ## [1.6.4] - 2022-06-01

pytorch_lightning/loops/dataloader/evaluation_loop.py

@@ -26,6 +26,7 @@
 from pytorch_lightning.accelerators import GPUAccelerator
 from pytorch_lightning.loops.dataloader import DataLoaderLoop
 from pytorch_lightning.loops.epoch import EvaluationEpochLoop
+from pytorch_lightning.loops.utilities import _set_sampler_epoch
 from pytorch_lightning.trainer.connectors.logger_connector.result import _OUT_DICT, _ResultCollection
 from pytorch_lightning.trainer.states import TrainerFn
 from pytorch_lightning.utilities.apply_func import apply_to_collection
@@ -161,14 +162,8 @@ def advance(self, *args: Any, **kwargs: Any) -> None:
             self._has_run = True
 
     def on_advance_start(self, *args: Any, **kwargs: Any) -> None:
-        dataloader = self.current_dataloader
-        if (
-            dataloader is not None
-            and getattr(dataloader, "sampler", None)
-            and callable(getattr(dataloader.sampler, "set_epoch", None))
-        ):
-            # set seed for distributed sampler (enables shuffling for each epoch)
-            dataloader.sampler.set_epoch(self.trainer.fit_loop.epoch_progress.current.processed)
+        if self.current_dataloader is not None:
+            _set_sampler_epoch(self.current_dataloader, self.trainer.fit_loop.epoch_progress.current.processed)
 
         super().on_advance_start(*args, **kwargs)
 

pytorch_lightning/loops/dataloader/prediction_loop.py

@@ -5,6 +5,7 @@
 
 from pytorch_lightning.loops.dataloader.dataloader_loop import DataLoaderLoop
 from pytorch_lightning.loops.epoch.prediction_epoch_loop import PredictionEpochLoop
+from pytorch_lightning.loops.utilities import _set_sampler_epoch
 from pytorch_lightning.strategies import DDPSpawnStrategy
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.types import _PREDICT_OUTPUT
@@ -87,13 +88,8 @@ def advance(self, *args: Any, **kwargs: Any) -> None:
         """Predicts one entire dataloader."""
         void(*args, **kwargs)
         dataloader = self.current_dataloader
-        if (
-            dataloader is not None
-            and getattr(dataloader, "sampler", None)
-            and callable(getattr(dataloader.sampler, "set_epoch", None))
-        ):
-            # set seed for distributed sampler (enables shuffling for each epoch)
-            dataloader.sampler.set_epoch(self.trainer.fit_loop.epoch_progress.current.processed)
+        if dataloader is not None:
+            _set_sampler_epoch(dataloader, self.trainer.fit_loop.epoch_progress.current.processed)
         dataloader = self.trainer.strategy.process_dataloader(dataloader)
         dataloader_iter = enumerate(dataloader)
         dl_max_batches = self.max_batches[self.current_dataloader_idx]

pytorch_lightning/loops/fit_loop.py

@@ -21,7 +21,7 @@
 from pytorch_lightning.loops import Loop
 from pytorch_lightning.loops.epoch import TrainingEpochLoop
 from pytorch_lightning.loops.epoch.training_epoch_loop import _OUTPUTS_TYPE as _EPOCH_OUTPUTS_TYPE
-from pytorch_lightning.loops.utilities import _is_max_limit_reached
+from pytorch_lightning.loops.utilities import _is_max_limit_reached, _set_sampler_epoch
 from pytorch_lightning.trainer.connectors.logger_connector.result import _ResultCollection
 from pytorch_lightning.trainer.progress import Progress
 from pytorch_lightning.trainer.supporters import TensorRunningAccum
@@ -232,11 +232,8 @@ def on_advance_start(self) -> None:  # type: ignore[override]
         # reset outputs here instead of in `reset` as they are not accumulated between epochs
         self._outputs = []
 
-        if self.trainer.train_dataloader is not None and callable(
-            getattr(self.trainer.train_dataloader.sampler, "set_epoch", None)
-        ):
-            # set seed for distributed sampler (enables shuffling for each epoch)
-            self.trainer.train_dataloader.sampler.set_epoch(self.epoch_progress.current.processed)
+        if self.trainer.train_dataloader is not None:
+            _set_sampler_epoch(self.trainer.train_dataloader, self.epoch_progress.current.processed)
 
         # changing gradient according accumulation_scheduler
         self.trainer.accumulation_scheduler.on_train_epoch_start(self.trainer, self.trainer.lightning_module)

pytorch_lightning/loops/utilities.py

@@ -21,6 +21,7 @@
 import numpy as np
 import torch
 from torch.optim import Optimizer
+from torch.utils.data import DataLoader
 
 import pytorch_lightning as pl
 from pytorch_lightning.loops import Loop
@@ -228,3 +229,16 @@ def _reset_progress(loop: Loop) -> None:
 def _v1_8_output_format(fx: Callable) -> bool:
     parameters = inspect.signature(fx).parameters
     return "new_format" in parameters and parameters["new_format"].default is True
+
+
+def _set_sampler_epoch(dataloader: DataLoader, epoch: int) -> None:
+    """Calls the ``set_epoch`` method on either the sampler or the batch sampler of the given dataloader.
+
+    Every PyTorch dataloader has either a sampler or a batch sampler, and if it is wrapped by a
+    :class:`~torch.utils.data.distributed.DistributedSampler`, ``set_epoch`` must be called at the beginning
+    of every epoch to ensure shuffling applies a new ordering. This has no effect if shuffling is off.
+    """
+    for sampler_name in ("sampler", "batch_sampler"):
+        sampler = getattr(dataloader, sampler_name, None)
+        if sampler is not None and callable(getattr(sampler, "set_epoch", None)):
+            sampler.set_epoch(epoch)

pytorch_lightning/trainer/supporters.py

@@ -438,9 +438,14 @@ class DataLoaderDict(dict):
 
     @property
     def sampler(self) -> Union[Iterable, Sequence, Mapping]:
-        """Return a collections of samplers extracting from loaders."""
+        """Return a collections of samplers extracted from loaders."""
         return apply_to_collection(self.loaders, (DataLoader, IterableDataset), getattr, "sampler", None)
 
+    @property
+    def batch_sampler(self) -> Union[Iterable, Sequence, Mapping]:
+        """Return a collections of batch samplers extracted from loaders."""
+        return apply_to_collection(self.loaders, (DataLoader, IterableDataset), getattr, "batch_sampler", None)
+
     def _wrap_loaders_max_size_cycle(self) -> Any:
         """Wraps all loaders to make sure they are cycled until the longest loader is exhausted.
 

tests/loops/test_evaluation_loop.py

@@ -12,11 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from unittest import mock
-from unittest.mock import Mock
+from unittest.mock import call, Mock
 
 import torch
 from torch.utils.data.dataloader import DataLoader
-from torch.utils.data.sampler import RandomSampler
+from torch.utils.data.sampler import BatchSampler, RandomSampler
 
 from pytorch_lightning import Trainer
 from pytorch_lightning.loops import EvaluationEpochLoop
@@ -44,9 +44,8 @@ def test_on_evaluation_epoch_end(eval_epoch_end_mock, tmpdir):
     assert eval_epoch_end_mock.call_count == 4
 
 
-def test_set_epoch_called_eval_predict(tmpdir):
-    """Tests that set_epoch (if the sampler has one) is called on the DataLoader during evaluation and
-    prediction."""
+def test_evaluation_loop_sampler_set_epoch_called(tmpdir):
+    """Tests that set_epoch is called on the dataloader's sampler (if any) during training and validation."""
 
     def _get_dataloader():
         dataset = RandomDataset(32, 64)
@@ -56,20 +55,60 @@ def _get_dataloader():
 
     model = BoringModel()
     trainer = Trainer(
-        default_root_dir=tmpdir, limit_train_batches=2, limit_val_batches=2, max_epochs=2, enable_model_summary=False
+        default_root_dir=tmpdir,
+        limit_train_batches=1,
+        limit_val_batches=1,
+        max_epochs=2,
+        enable_model_summary=False,
+        enable_checkpointing=False,
+        logger=False,
+    )
+
+    train_dataloader = _get_dataloader()
+    val_dataloader = _get_dataloader()
+    trainer.fit(model, train_dataloaders=train_dataloader, val_dataloaders=val_dataloader)
+    # One for each epoch
+    assert train_dataloader.sampler.set_epoch.call_args_list == [call(0), call(1)]
+    # One for each epoch + sanity check
+    assert val_dataloader.sampler.set_epoch.call_args_list == [call(0), call(0), call(1)]
+
+    val_dataloader = _get_dataloader()
+    trainer.validate(model, val_dataloader)
+    assert val_dataloader.sampler.set_epoch.call_args_list == [call(2)]
+
+
+def test_evaluation_loop_batch_sampler_set_epoch_called(tmpdir):
+    """Tests that set_epoch is called on the dataloader's batch sampler (if any) during training and validation."""
+
+    def _get_dataloader():
+        dataset = RandomDataset(32, 64)
+        sampler = RandomSampler(dataset)
+        batch_sampler = BatchSampler(sampler, 2, True)
+        batch_sampler.set_epoch = Mock()
+        return DataLoader(dataset, batch_sampler=batch_sampler)
+
+    model = BoringModel()
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        limit_train_batches=1,
+        limit_val_batches=1,
+        max_epochs=2,
+        enable_model_summary=False,
+        enable_checkpointing=False,
+        logger=False,
     )
 
     train_dataloader = _get_dataloader()
     val_dataloader = _get_dataloader()
     trainer.fit(model, train_dataloaders=train_dataloader, val_dataloaders=val_dataloader)
     # One for each epoch
-    assert train_dataloader.sampler.set_epoch.call_count == 2
+    assert train_dataloader.batch_sampler.set_epoch.call_args_list == [call(0), call(1)]
     # One for each epoch + sanity check
-    assert val_dataloader.sampler.set_epoch.call_count == 3
+    assert val_dataloader.batch_sampler.set_epoch.call_args_list == [call(0), call(0), call(1)]
 
     val_dataloader = _get_dataloader()
     trainer.validate(model, val_dataloader)
-    assert val_dataloader.sampler.set_epoch.call_count == 1
+    assert val_dataloader.batch_sampler.set_epoch.call_args_list == [call(2)]
 
 
 @mock.patch(

tests/loops/test_utilities.py

@@ -11,10 +11,12 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from unittest.mock import Mock
+
 import pytest
 import torch
 
-from pytorch_lightning.loops.utilities import _extract_hiddens, _v1_8_output_format
+from pytorch_lightning.loops.utilities import _extract_hiddens, _set_sampler_epoch, _v1_8_output_format
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 
 
@@ -61,3 +63,23 @@ def training_epoch_end(outputs, new_format=True):
         ...
 
     assert _v1_8_output_format(training_epoch_end)
+
+
+def test_set_sampler_epoch():
+    # No samplers
+    dataloader = Mock()
+    dataloader.sampler = None
+    dataloader.batch_sampler = None
+    _set_sampler_epoch(dataloader, 55)
+
+    # set_epoch not callable
+    dataloader = Mock()
+    dataloader.sampler.set_epoch = None
+    dataloader.batch_sampler.set_epoch = None
+    _set_sampler_epoch(dataloader, 55)
+
+    # set_epoch callable
+    dataloader = Mock()
+    _set_sampler_epoch(dataloader, 55)
+    dataloader.sampler.set_epoch.assert_called_once_with(55)
+    dataloader.batch_sampler.set_epoch.assert_called_once_with(55)