diff --git a/pytorch_lightning/trainer/training_loop.py b/pytorch_lightning/trainer/training_loop.py
index 48bb359713..7c8d0143b8 100644
--- a/pytorch_lightning/trainer/training_loop.py
+++ b/pytorch_lightning/trainer/training_loop.py
@@ -151,7 +151,7 @@ When this flag is enabled each batch is split into sequences of size truncated_b
 
 
 """
-
+import copy
 import inspect
 from abc import ABC, abstractmethod
 import warnings
@@ -586,7 +586,7 @@ class TrainerTrainLoopMixin(ABC):
             gpu_id = 0
             if isinstance(self.data_parallel_device_ids, list):
                 gpu_id = self.data_parallel_device_ids[0]
-            batch = self.transfer_batch_to_gpu(batch.copy(), gpu_id)
+            batch = self.transfer_batch_to_gpu(copy.copy(batch), gpu_id)
             args[0] = batch
             output = self.model.training_step(*args)