Enable distributed training with CombinedDataLoader and max_size_cycle (#10374)

Co-authored-by: Carlos Mocholi <[email protected]>
Co-authored-by: Thomas Chaton <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -9,7 +9,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
--
+- Fixed `CombinedLoader` and `max_size_cycle` didn't receive a `DistributedSampler` ([#10374](https://github.com/PyTorchLightning/pytorch-lightning/issues/10374))
+
 
 
 ## [1.5.1] - 2021-11-09

pytorch_lightning/trainer/data_loading.py

@@ -28,7 +28,7 @@
 from pytorch_lightning.overrides.distributed import IndexBatchSamplerWrapper, UnrepeatedDistributedSampler
 from pytorch_lightning.trainer.connectors.accelerator_connector import AcceleratorConnector
 from pytorch_lightning.trainer.states import RunningStage
-from pytorch_lightning.trainer.supporters import CombinedLoader
+from pytorch_lightning.trainer.supporters import CombinedLoader, CycleIterator
 from pytorch_lightning.utilities import rank_zero_warn
 from pytorch_lightning.utilities.apply_func import apply_to_collection
 from pytorch_lightning.utilities.auto_restart import (
@@ -136,14 +136,22 @@ def prepare_dataloader(self, dataloader: Any, shuffle: bool, mode: Optional[Runn
         if isinstance(dataloader, CombinedLoader):
             # apply `prepare_dataloader` on all the collection of loaders
             dataloader.loaders = apply_to_collection(
-                dataloader.loaders, DataLoader, self.prepare_dataloader, shuffle, mode=mode
+                dataloader.loaders, (DataLoader, CycleIterator), self.prepare_dataloader, shuffle, mode=mode
             )
+            # the length need to recomputed across all dataloaders in case of special behavior.
+            dataloader._apply_cycle_iterator_length()
             return dataloader
 
         # don't do anything if it's not a dataloader
-        if not isinstance(dataloader, DataLoader):
+        if not isinstance(dataloader, (DataLoader, CycleIterator)):
             return dataloader
 
+        cycle_iterator: Optional[CycleIterator] = None
+
+        if isinstance(dataloader, CycleIterator):
+            cycle_iterator = dataloader
+            dataloader = dataloader.loader
+
         if (
             _fault_tolerant_training()  # injects components to track the state
             or self._requires_distributed_sampler(dataloader)  # sets the distributed sampler
@@ -153,6 +161,10 @@ def prepare_dataloader(self, dataloader: Any, shuffle: bool, mode: Optional[Runn
             sampler = self._resolve_sampler(dataloader, shuffle=shuffle, mode=mode)
             dataloader = self._update_dataloader(dataloader, sampler, mode=mode)
 
+        if cycle_iterator is not None:
+            cycle_iterator.loader = dataloader
+            return cycle_iterator
+
         return dataloader
 
     def _resolve_sampler(self, dataloader: DataLoader, shuffle: bool, mode: Optional[RunningStage] = None) -> Sampler:

pytorch_lightning/trainer/supporters.py

@@ -457,6 +457,19 @@ def _wrap_loaders_max_size_cycle(self) -> Any:
             )
             state.reset()
 
+    def _apply_cycle_iterator_length(self) -> None:
+        """When the model is `max_size_cycle`, compute the length across all ``CycleIterator`` and re-assign it to
+        all dataloaders."""
+        if self.mode != "max_size_cycle":
+            return
+
+        def set_len(cycle_iterator: CycleIterator, length: int) -> None:
+            cycle_iterator.length = length
+
+        all_lengths = apply_to_collection(self.loaders, CycleIterator, lambda c: get_len(c.loader))
+        max_length = _nested_calc_num_data(all_lengths, max)
+        apply_to_collection(self.loaders, CycleIterator, set_len, length=max_length)
+
     def __iter__(self) -> Any:
         """Create and return an iterator, `CombinedLoaderIterator`, for the combined loader."""
 
@@ -473,11 +486,12 @@ def __getstate__patch__(*_):
         return iterator
 
     @staticmethod
-    def _calc_num_batches(loaders: Any) -> Union[int, float]:
+    def _calc_num_batches(loaders: Any, mode="min_size") -> Union[int, float]:
         """Compute the length (aka the number of batches) of `CombinedLoader`.
 
         Args:
             loaders: a collections of loaders.
+            mode: Mode used by the CombinedDataloader
 
         Returns:
             length: the minimum length of loaders
@@ -486,10 +500,10 @@ def _calc_num_batches(loaders: Any) -> Union[int, float]:
 
         if isinstance(all_lengths, (int, float)):
             return all_lengths
-        return _nested_calc_num_data(all_lengths, min)
+        return _nested_calc_num_data(all_lengths, max if mode == "max_size_cycle" else min)
 
     def __len__(self) -> int:
-        return self._calc_num_batches(self.loaders)
+        return self._calc_num_batches(self.loaders, mode=self.mode)
 
     @staticmethod
     def _shutdown_workers_and_reset_iterator(dataloader) -> None:

tests/trainer/test_supporters.py

@@ -33,8 +33,10 @@
 )
 from pytorch_lightning.utilities.apply_func import apply_to_collection
 from pytorch_lightning.utilities.auto_restart import CaptureMapDataset, FastForwardSampler
+from pytorch_lightning.utilities.data import get_len
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.imports import _TORCH_GREATER_EQUAL_1_7
+from tests.helpers.boring_model import RandomDataset
 
 
 def test_tensor_running_accum_reset():
@@ -379,3 +381,56 @@ def _assert_dataset(loader):
             assert isinstance(d, CustomDataset)
 
     apply_to_collection(dataloader.loaders, DataLoader, _assert_dataset)
+
+
+@pytest.mark.parametrize("replace_sampler_ddp", [False, True])
+def test_combined_data_loader_with_max_size_cycle_and_ddp(replace_sampler_ddp, tmpdir):
+    """This test makes sure distributed sampler has been properly injected in dataloaders when using CombinedLoader
+    with ddp and `max_size_cycle` mode."""
+    trainer = Trainer(strategy="ddp", accelerator="auto", devices=2, replace_sampler_ddp=replace_sampler_ddp)
+
+    dataloader = CombinedLoader(
+        {"a": DataLoader(RandomDataset(32, 8), batch_size=1), "b": DataLoader(RandomDataset(32, 8), batch_size=1)},
+    )
+    dataloader = trainer.prepare_dataloader(dataloader, shuffle=False)
+    assert len(dataloader) == 4 if replace_sampler_ddp else 8
+
+    for a_length in [6, 8, 10]:
+        dataloader = CombinedLoader(
+            {
+                "a": DataLoader(range(a_length), batch_size=1),
+                "b": DataLoader(range(8), batch_size=1),
+            },
+            mode="max_size_cycle",
+        )
+
+        length = max(a_length, 8)
+        assert len(dataloader) == length
+        dataloader = trainer.prepare_dataloader(dataloader, shuffle=False)
+        assert len(dataloader) == length // 2 if replace_sampler_ddp else length
+        if replace_sampler_ddp:
+            last_batch = list(dataloader)[-1]
+            if a_length == 6:
+                assert last_batch == {"a": torch.tensor([0]), "b": torch.tensor([6])}
+            elif a_length == 8:
+                assert last_batch == {"a": torch.tensor([6]), "b": torch.tensor([6])}
+            elif a_length == 10:
+                assert last_batch == {"a": torch.tensor([8]), "b": torch.tensor([0])}
+
+    class InfiniteDataset(IterableDataset):
+        def __iter__(self):
+            while True:
+                yield 1
+
+    dataloader = CombinedLoader(
+        {
+            "a": DataLoader(InfiniteDataset(), batch_size=1),
+            "b": DataLoader(range(8), batch_size=1),
+        },
+        mode="max_size_cycle",
+    )
+    assert get_len(dataloader) == float("inf")
+    assert len(dataloader.loaders["b"].loader) == 8
+    dataloader = trainer.prepare_dataloader(dataloader, shuffle=False)
+    assert len(dataloader.loaders["b"].loader) == 4 if replace_sampler_ddp else 8
+    assert get_len(dataloader) == float("inf")