feat: save checkpoint before deleting old ones (#1453)

* feat: save checkpoint before deleting old ones

* fix: make sure that the new model is not deleted

* changelog

Co-authored-by: J. Borovec <jirka.borovec@seznam.cz>
Co-authored-by: William Falcon <waf2107@columbia.edu>

CHANGELOG.md

@@ -34,6 +34,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+
+- Fixed saving checkpoint before deleting old ones ([#1453](https://github.com/PyTorchLightning/pytorch-lightning/pull/1453))
+
 - Fixed loggers - flushing last logged metrics even before continue, e.g. `trainer.test()` results ([#1459](https://github.com/PyTorchLightning/pytorch-lightning/pull/1459))
 
 - Added a missing call to the `on_before_zero_grad` model hook ([#1493](https://github.com/PyTorchLightning/pytorch-lightning/pull/1493)). 

pytorch_lightning/callbacks/model_checkpoint.py

@@ -216,10 +216,12 @@ class ModelCheckpoint(Callback):
 
     def _do_check_save(self, filepath, current, epoch):
         # remove kth
+
+        del_list = []
         if len(self.best_k_models) == self.save_top_k and self.save_top_k > 0:
             delpath = self.kth_best_model
             self.best_k_models.pop(self.kth_best_model)
-            self._del_model(delpath)
+            del_list.append(delpath)
 
         self.best_k_models[filepath] = current
         if len(self.best_k_models) == self.save_top_k:
@@ -238,3 +240,7 @@ class ModelCheckpoint(Callback):
                 f' {current:0.5f} (best {self.best:0.5f}), saving model to'
                 f' {filepath} as top {self.save_top_k}')
         self._save_model(filepath)
+
+        for cur_path in del_list:
+            if cur_path != filepath:
+                self._del_model(cur_path)