diff --git a/train.py b/train.py
index 1a6a98c5..b87600d5 100644
--- a/train.py
+++ b/train.py
@@ -180,15 +180,16 @@ def train(
                               model) is nn.parallel.DistributedDataParallel else model.state_dict(),
                           'optimizer': optimizer.state_dict()}
             torch.save(checkpoint, latest)
-            del checkpoint
 
             # Save best checkpoint
             if best_loss == mloss['total']:
-                os.system('cp ' + latest + ' ' + best)
+                torch.save(checkpoint, best)
 
             # Save backup weights every 10 epochs (optional)
             if epoch > 0 and epoch % 10 == 0:
-                os.system('cp ' + latest + ' ' + weights + 'backup%g.pt' % epoch)
+                torch.save(checkpoint, weights + 'backup%g.pt' % epoch)
+
+            del checkpoint
 
         # Calculate mAP
         with torch.no_grad():