Issue 1986 Checkpointing write & remove order change

ignite/contrib/handlers/clearml_logger.py

@@ -806,7 +806,11 @@ def get_local_copy(self, filename: str) -> Optional[str]:
 
     @idist.one_rank_only()
     def remove(self, filename: str) -> None:
-        super(ClearMLSaver, self).remove(filename)
+        from clearml.storage.helper import StorageHelper
+
+        clearml_logger = self._task.get_logger()
+        helper = StorageHelper.get(filename, logger=clearml_logger)
+        helper.delete(filename)
         for slots in self._checkpoint_slots.values():
             try:
                 slots[slots.index(filename)] = None

ignite/contrib/handlers/neptune_logger.py

@@ -78,7 +78,7 @@ class NeptuneLogger(BaseLogger):
                 project_name="shared/pytorch-ignite-integration",
                 experiment_name="cnn-mnist", # Optional,
                 params={"max_epochs": 10}, # Optional,
-                tags=["pytorch-ignite","minst"] # Optional
+                tags=["pytorch-ignite","mnist"] # Optional
             )
 
             # Attach the logger to the trainer to log training loss at each iteration

ignite/handlers/checkpoint.py

@@ -363,7 +363,7 @@ def __call__(self, engine: Engine) -> None:
                 global_step = engine.state.get_event_attrib_value(Events.ITERATION_COMPLETED)
             priority = global_step
 
-        if self._check_lt_n_saved() or self._compare_fn(priority):
+        if self._check_lt_n_saved(or_equal=True) or self._compare_fn(priority):
 
             priority_str = f"{priority}" if isinstance(priority, numbers.Integral) else f"{priority:.4f}"
 
@@ -401,18 +401,6 @@ def __call__(self, engine: Engine) -> None:
                 "priority": priority,
             }
 
-            try:
-                index = list(map(lambda it: it.filename == filename, self._saved)).index(True)
-                to_remove = True
-            except ValueError:
-                index = 0
-                to_remove = not self._check_lt_n_saved()
-
-            if to_remove:
-                item = self._saved.pop(index)
-                if isinstance(self.save_handler, BaseSaveHandler):
-                    self.save_handler.remove(item.filename)
-
             self._saved.append(Checkpoint.Item(priority, filename))
             self._saved.sort(key=lambda it: it[0])
 
@@ -425,6 +413,14 @@ def __call__(self, engine: Engine) -> None:
             except TypeError:
                 self.save_handler(checkpoint, filename)
 
+            index = list(map(lambda it: it.filename == filename, self._saved)).index(True)
+            to_remove = not self._check_lt_n_saved(or_equal=True)
+
+            if to_remove:
+                item = self._saved.pop(index)
+                if isinstance(self.save_handler, BaseSaveHandler):
+                    self.save_handler.remove(item.filename)
+
     def _setup_checkpoint(self) -> Dict[str, Dict[Any, Any]]:
         checkpoint = {}
         if self.to_save is not None:
@@ -630,14 +626,18 @@ class DiskSaver(BaseSaveHandler):
         dirname: Directory path where the checkpoint will be saved
         atomic: if True, checkpoint is serialized to a temporary file, and then
             moved to final destination, so that files are guaranteed to not be damaged
-            (for example if exception occurs during saving).
+            (for example if exception occurs during saving). Setting ``atomic=True`` is
+            recommended if ``n_saved=1`` is set in checkpoint object. For details, see
+            issue `#1986`_
         create_dir: if True, will create directory ``dirname`` if it doesnt exist.
         require_empty: If True, will raise exception if there are any files in the
             directory ``dirname``.
         kwargs: Accepted keyword arguments for `torch.save` or `xm.save`.
 
     .. versionchanged:: 0.4.2
         Accept ``kwargs`` for `torch.save` or `xm.save`.
+
+    .. _#1986: https://github.com/pytorch/ignite/issues/1986
     """
 
     def __init__(