fix checkpoint dir issue under distributed training

tensorpack/callbacks/saver.py

@@ -42,7 +42,10 @@ class ModelSaver(Callback):
         if checkpoint_dir is not None:
             if not tf.gfile.IsDirectory(checkpoint_dir):  # v2: tf.io.gfile.isdir
                 tf.gfile.MakeDirs(checkpoint_dir)  # v2: tf.io.gfile.makedirs
-        self.checkpoint_dir = os.path.normpath(checkpoint_dir)
+        # If None, allow it to be init, but fail later if used
+        # For example, if chief_only=True, it can still be safely initialized
+        # in non-chief workers which don't have logger dir
+        self.checkpoint_dir = os.path.normpath(checkpoint_dir) if checkpoint_dir is not None else checkpoint_dir
 
     def _setup_graph(self):
         assert self.checkpoint_dir is not None, \