Correct assignement for logits in classifier example

I tried to address https://github.com/huggingface/pytorch-pretrained-BERT/issues/76
should be correct, but there's likely a more efficient way.

examples/run_classifier.py

@@ -605,7 +605,8 @@ def main():
             label_ids = label_ids.to(device)
 
             with torch.no_grad():
-                tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids)
+                tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids)
+                logits = model(input_ids, segment_ids, input_mask)
 
             logits = logits.detach().cpu().numpy()
             label_ids = label_ids.to('cpu').numpy()