comments

Author: ananthsub

pytorch_lightning/trainer/training_loop.py

@@ -199,6 +199,12 @@ def on_train_batch_end(self, epoch_output, batch_end_outputs, batch, batch_idx,
         self.trainer.logger_connector.on_train_batch_end()
 
     def reset_train_val_dataloaders(self, model):
+        """
+        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)
 

tests/trainer/test_dataloaders.py

@@ -1112,6 +1112,14 @@ def test_dataloaders_load_only_once_val_interval(tmpdir):
     expected_sequence = [
         'val_dataloader',
         'train_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 datalodaer reloaded.
+        # This breaks the assumption behind reload_dataloaders_every_epoch=True
         'val_dataloader',
         'val_dataloader',
         'val_dataloader',