Improved memory efficiency

bounded_attention.py

@@ -301,7 +301,7 @@ class BoundedAttention(injection_utils.AttentionBase):
 
         mask_maps.append(attn.mean(dim=1))  # mean over heads
         if self.cur_step > 0:
-            mask_maps = mask_maps[1:]  # throw away old maps
+            mask_maps.pop(0)  # throw away old maps
 
     def _save_loss_values(self, attn, is_cross):
         if (