Raise an error when lightning replaces an existing sampler (#2020)

* Raise an error when lightning replaces an existing sampler

Currently, Trainer replaces the existing sampler with DistributedSampler
if running distributing training and `replace_sampler_ddp=True` (default
behaviour). If a user has configured an existing sampler, this would
lead to widely different results if running a distributed vs
non-distributed training.

This PR fixes this by raising an Error if user has configured a sampler
and uses `replace_sampler_ddp=True`. The recommended behavior from now
on is to either remove the sampler or set `replace_sampler_ddp=False`

* Fix tests

* Simpler fix

* Fix tests

* Make inner method protected

* Apply suggestions from code review

Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>

pytorch_lightning/trainer/data_loading.py

@@ -3,7 +3,7 @@ from abc import ABC, abstractmethod
 from typing import Union, List, Tuple, Callable
 
 import torch.distributed as torch_distrib
-from torch.utils.data import DataLoader, RandomSampler
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from torch.utils.data.distributed import DistributedSampler
 
 from pytorch_lightning.core import LightningModule
@@ -113,39 +113,39 @@ class TrainerDataLoadingMixin(ABC):
         need_dist_sampler = (self.use_ddp or self.use_ddp2 or self.use_horovod or self.use_tpu)
 
         if self.replace_sampler_ddp and need_dist_sampler:
+            if not isinstance(dataloader.sampler, (SequentialSampler, RandomSampler)):
+                raise MisconfigurationException(
+                    'You seem to have configured a sampler in your DataLoader. This will be replaced '
+                    ' by `DistributedSampler` since `replace_sampler_ddp` is True and you are using'
+                    ' distributed training. Either remove the sampler from your DataLoader or set'
+                    ' `replace_sampler_ddp`=False if you want to use your custom sampler.')
+
             skip_keys = ['sampler', 'batch_sampler', 'dataset_kind']
 
             dl_args = {
                 k: v for k, v in dataloader.__dict__.items() if not k.startswith('_') and k not in skip_keys
             }
 
-            if self.use_tpu:
-                sampler = DistributedSampler(
-                    dataloader.dataset,
-                    num_replicas=xm.xrt_world_size(),
-                    rank=xm.get_ordinal(),
-                )
-            elif self.use_horovod:
-                sampler = DistributedSampler(dataloader.dataset,
-                                             num_replicas=hvd.size(),
-                                             rank=hvd.rank())
-            else:
-                world_size = {
-                    'ddp': self.num_nodes * self.num_processes,
-                    'ddp2': self.num_nodes,
-                    'ddp_cpu': self.num_processes * self.num_nodes
-                }
-                sampler = DistributedSampler(
-                    dataloader.dataset,
-                    num_replicas=world_size[self.distributed_backend],
-                    rank=self.proc_rank,
-                )
-
-            dl_args['sampler'] = sampler
+            dl_args['sampler'] = self._get_distributed_sampler(dataloader)
             dataloader = type(dataloader)(**dl_args)
 
         return dataloader
 
+    def _get_distributed_sampler(self, dataloader):
+        if self.use_tpu:
+            kwargs = dict(num_replicas=xm.xrt_world_size(), rank=xm.get_ordinal())
+        elif self.use_horovod:
+            kwargs = dict(num_replicas=hvd.size(), rank=hvd.rank())
+        else:
+            world_size = {
+                'ddp': self.num_nodes * self.num_processes,
+                'ddp2': self.num_nodes,
+                'ddp_cpu': self.num_processes * self.num_nodes
+            }
+            kwargs = dict(num_replicas=world_size[self.distributed_backend], rank=self.proc_rank)
+        sampler = DistributedSampler(dataloader.dataset, **kwargs)
+        return sampler
+
     def reset_train_dataloader(self, model: LightningModule) -> None:
         """Resets the train dataloader and initialises required variables
         (number of batches, when to validate, etc.).

tests/trainer/test_dataloaders.py

@@ -416,6 +416,20 @@ def test_dataloader_reinit_for_subclass():
     assert isinstance(result, CustomDataLoader)
     assert hasattr(result, 'dummy_kwarg')
 
+    # Shuffled DataLoader should also work
+    result = trainer.auto_add_sampler(CustomDataLoader(list(range(1000)), shuffle=True), train=True)
+    assert isinstance(result, torch.utils.data.DataLoader)
+    assert isinstance(result, CustomDataLoader)
+    assert hasattr(result, 'dummy_kwarg')
+
+    class CustomSampler(torch.utils.data.Sampler):
+        pass
+
+    # Should raise an error if existing sampler is being replaced
+    with pytest.raises(MisconfigurationException, match='DistributedSampler'):
+        trainer.auto_add_sampler(
+            CustomDataLoader(list(range(1000)), sampler=CustomSampler(list(range(1000)))), train=True)
+
 
 @pytest.mark.skipif(torch.cuda.device_count() < 3, reason='Test requires multiple GPUs')
 def test_batch_size_smaller_than_num_gpus():