refactored dataloader process hook (#3139)

pytorch_lightning/accelerators/base_backend.py

@@ -22,3 +22,6 @@ class Accelerator(object):
 
     def validation_step_end(self, output):
         return output
+
+    def process_dataloader(self, dataloader):
+        return dataloader

pytorch_lightning/accelerators/tpu_backend.py

@@ -27,6 +27,7 @@ try:
     import torch_xla
     import torch_xla.core.xla_model as xm
     import torch_xla.distributed.xla_multiprocessing as xmp
+    import torch_xla.distributed.parallel_loader as xla_pl
 except ImportError:
     XLA_AVAILABLE = False
 else:
@@ -139,6 +140,12 @@ class TPUBackend(Accelerator):
         output = self.trainer.model.test_step(*args)
         return output
 
+    def process_dataloader(self, dataloader):
+        device = xm.xla_device(self.trainer.tpu_id)
+        dataloader = xla_pl.ParallelLoader(dataloader, [device])
+        dataloader = dataloader.per_device_loader(device)
+        return dataloader
+
     def to_device(self, batch):
         """
         Transfers the data to the TPU.

pytorch_lightning/trainer/evaluation_loop.py

@@ -260,10 +260,7 @@ class TrainerEvaluationLoopMixin(ABC):
             dl_outputs = []
 
             # on TPU we have to wrap it under the ParallelLoader
-            if self.use_tpu:
+            dataloader = self.accelerator_backend.process_dataloader(dataloader)
-                device = xm.xla_device(self.tpu_id)
-                dataloader = xla_pl.ParallelLoader(dataloader, [device])
-                dataloader = dataloader.per_device_loader(device)
 
             # each dataloader has a max num batches
             dl_max_batches = self.evaluation_loop.max_batches[dataloader_idx]

pytorch_lightning/trainer/training_loop.py

@@ -427,15 +427,6 @@ class TrainerTrainLoopMixin(ABC):
 
                 self.run_training_teardown()
 
-    def prepare_train_loop_dataloader(self, train_dataloader):
-        # on TPU we have to wrap it under the ParallelLoader
-        if self.use_tpu:
-            device = xm.xla_device(self.tpu_id)
-            train_dataloader = xla_pl.ParallelLoader(train_dataloader, [device])
-            train_dataloader = train_dataloader.per_device_loader(device)
-
-        return train_dataloader
-
     def run_on_epoch_start_hook(self, model):
         # Epoch start events
         with self.profiler.profile('on_epoch_start'):
@@ -464,7 +455,7 @@ class TrainerTrainLoopMixin(ABC):
         self.run_on_epoch_start_hook(model)
 
         # modify dataloader if needed (ddp, etc...)
-        train_dataloader = self.prepare_train_loop_dataloader(self.train_dataloader)
+        train_dataloader = self.accelerator_backend.process_dataloader(self.train_dataloader)
 
         # bookkeeping
         num_optimizers = len(self._get_optimizers_iterable())