diff --git a/pytorch_lightning/trainer/train_loop_mixin.py b/pytorch_lightning/trainer/train_loop_mixin.py
index 7e9fde01485e3..6130fd4931342 100644
--- a/pytorch_lightning/trainer/train_loop_mixin.py
+++ b/pytorch_lightning/trainer/train_loop_mixin.py
@@ -290,7 +290,7 @@ def training_forward(self, batch, batch_nb, opt_idx, hiddens):
             gpu_id = 0
             if type(self.data_parallel_device_ids) is list:
                 gpu_id = self.data_parallel_device_ids[0]
-            batch = self.transfer_batch_to_gpu(batch, gpu_id)
+            batch = self.transfer_batch_to_gpu(batch.copy(), gpu_id)
             args[0] = batch
             output = self.model.training_step(*args)