improvements

Lightning-AI · Jan 19, 2022 · 0e0ebd8 · 0e0ebd8
1 parent 5e2b98c
commit 0e0ebd8
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 10 deletions.
diff --git a/_notebooks b/_notebooks
diff --git a/docs/source/common/test_set.rst b/docs/source/common/test_set.rst
@@ -50,7 +50,7 @@ To run the test set after training completes, use this method.
 
 .. warning::
 
-    It is recommended to test on single device since distributed strategies such as DDP
+    It is recommended to test with ``Trainer(devices=1)`` since distributed strategies such as DDP
     uses :class:`~torch.utils.data.distributed.DistributedSampler` internally, which replicates some samples to
     make sure all devices have same batch size in case of uneven inputs. This is helpful to make sure
     benchmarking for research papers is done the right way.
@@ -146,7 +146,7 @@ Apart from this ``.validate`` has same API as ``.test``, but would rely respecti
 
 .. warning::
 
-    When using ``trainer.validate()``, it is recommended to use a single device since distributed strategies such as DDP
+    When using ``trainer.validate()``, it is recommended to use ``Trainer(devices=1)`` since distributed strategies such as DDP
     uses :class:`~torch.utils.data.distributed.DistributedSampler` internally, which replicates some samples to
     make sure all devices have same batch size in case of uneven inputs. This is helpful to make sure
     benchmarking for research papers is done the right way.

diff --git a/pytorch_lightning/trainer/connectors/data_connector.py b/pytorch_lightning/trainer/connectors/data_connector.py
@@ -390,13 +390,14 @@ def _resolve_sampler(self, dataloader: DataLoader, shuffle: bool, mode: Optional
                 **self.trainer.distributed_sampler_kwargs,
             )
 
+            # update docs too once this is resolved
             trainer_fn = self.trainer.state.fn
             if isinstance(sampler, DistributedSampler) and trainer_fn in (TrainerFn.VALIDATING, TrainerFn.TESTING):
                 rank_zero_warn(
                     f"Using `DistributedSampler` with the dataloaders. During trainer.{trainer_fn.value}(),"
-                    " it is recommended to use single device strategy to ensure each sample/batch gets evaluated"
-                    " exactly once. Otherwise multi-device setting uses `DistributedSampler` replicates some samples"
-                    " to make sure all devices have same batch size in case of uneven inputs."
+                    " it is recommended to use `Trainer(devices=1)` to ensure each sample/batch gets evaluated"
+                    " exactly once. Otherwise, multi-device settings use `DistributedSampler` that replicates"
+                    " some samples to make sure all devices have same batch size in case of uneven inputs."
                 )
 
             return sampler

diff --git a/tests/trainer/connectors/test_data_connector.py b/tests/trainer/connectors/test_data_connector.py
@@ -76,10 +76,10 @@ def test_eval_distributed_sampler_warning(tmpdir):
     trainer = Trainer(strategy="ddp", devices=2, accelerator="cpu", fast_dev_run=True)
     trainer._data_connector.attach_data(model)
 
-    with pytest.warns(UserWarning, match="use single device strategy to ensure each sample"):
-        trainer.state.fn = TrainerFn.VALIDATING
+    trainer.state.fn = TrainerFn.VALIDATING
+    with pytest.warns(UserWarning, match="multi-device settings use `DistributedSampler`"):
         trainer.reset_val_dataloader(model)
 
-    with pytest.warns(UserWarning, match="use single device strategy to ensure each sample"):
-        trainer.state.fn = TrainerFn.TESTING
+    trainer.state.fn = TrainerFn.TESTING
+    with pytest.warns(UserWarning, match="multi-device settings use `DistributedSampler`"):
         trainer.reset_test_dataloader(model)