[fix] Attach train+val dataloaders to trainer in trainer loop (#7207)

* Update training_loop.py

* Update test_dataloaders.py

* changelog

* delay reload

* go back

* comments

* Update training_loop.py

* Update test_dataloaders.py

* Update tests/trainer/test_dataloaders.py

Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>

CHANGELOG.md

@@ -290,6 +290,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+
+- Fixed attaching train and validation dataloaders when `reload_dataloaders_every_epoch=True` and `num_sanity_val_steps=0` ([#7207](https://github.com/PyTorchLightning/pytorch-lightning/pull/7207))
+
+
 - Added a barrier in the accelerator `teardown` to synchronize processes before execution finishes ([#6814](https://github.com/PyTorchLightning/pytorch-lightning/pull/6814))
 
 

pytorch_lightning/trainer/training_loop.py

@@ -174,11 +174,17 @@ class TrainLoop:
         # reset batch logger internals
         self.trainer.logger_connector.on_train_batch_end()
 
-    def reset_train_val_dataloaders(self, model):
-        if self.trainer.train_dataloader is None or not self.trainer.reload_dataloaders_every_epoch:
+    def reset_train_val_dataloaders(self, model) -> None:
+        """
+        Resets train and val dataloaders if none are attached to the trainer.
+
+        The val dataloader must be initialized before training loop starts, as the training loop
+        inspects the val dataloader to determine whether to run the evaluation loop.
+        """
+        if self.trainer.train_dataloader is None:
             self.trainer.reset_train_dataloader(model)
 
-        if self.trainer.val_dataloaders is None and not self.trainer.reload_dataloaders_every_epoch:
+        if self.trainer.val_dataloaders is None:
             self.trainer.reset_val_dataloader(model)
 
     def track_epoch_end_reduce_metrics(self, epoch_output, batch_end_outputs):

tests/trainer/test_dataloaders.py

@@ -25,6 +25,7 @@ from torch.utils.data.sampler import SequentialSampler
 
 import tests.helpers.pipelines as tpipes
 from pytorch_lightning import Callback, seed_everything, Trainer
+from pytorch_lightning.callbacks import ModelCheckpoint
 from pytorch_lightning.trainer.states import TrainerState
 from pytorch_lightning.utilities import _TORCH_GREATER_EQUAL_1_6
 from pytorch_lightning.utilities.data import has_iterable_dataset, has_len
@@ -1234,7 +1235,16 @@ def test_dataloaders_load_every_epoch(tmpdir):
 @mock.patch.dict(os.environ, {"PL_DEV_DEBUG": "1"})
 def test_dataloaders_load_every_epoch_no_sanity_check(tmpdir):
 
-    model = EvalModelTemplate()
+    class TestModel(BoringModel):
+
+        def validation_step(self, batch, batch_idx):
+            self.log("dummy_val", 5.0)
+            return super().validation_step(batch, batch_idx)
+
+    model = TestModel()
+
+    # This callback tests that the evaluation metrics are available by the time we run checkpointing
+    checkpoint_callback = ModelCheckpoint(monitor="dummy_val", save_top_k=1)
 
     # logger file to get meta
     trainer = Trainer(
@@ -1244,21 +1254,32 @@ def test_dataloaders_load_every_epoch_no_sanity_check(tmpdir):
         num_sanity_val_steps=0,
         reload_dataloaders_every_epoch=True,
         max_epochs=3,
+        callbacks=[checkpoint_callback],
     )
     trainer.fit(model)
     assert trainer.state == TrainerState.FINISHED, f"Training failed with {trainer.state}"
 
     trainer.test()
 
-    assert len(trainer.dev_debugger.val_dataloader_calls) == 3
+    assert len(trainer.dev_debugger.val_dataloader_calls) == 4
     assert len(trainer.dev_debugger.train_dataloader_calls) == 3
     assert len(trainer.dev_debugger.test_dataloader_calls) == 1
 
     # verify the sequence
     calls = trainer.dev_debugger.dataloader_sequence_calls
+
     expected_sequence = [
         'train_dataloader',
         'val_dataloader',
+        # This has subsequent calls to val_dataloader
+        # because the training loop runs the evaluation loop,
+        # which reloads the val dataloader again.
+        # We cannot yet rely on trainer.current_epoch=0 to skip reloading
+        # the val dataloader on the first epoch because this only tracks the training epoch
+        # meaning multiple passes through the validation data within a single training epoch
+        # would not have the dataloader reloaded.
+        # This breaks the assumption behind reload_dataloaders_every_epoch=True
+        'val_dataloader',
         'train_dataloader',
         'val_dataloader',
         'train_dataloader',