Not always consider a local model a checkpoint in run_glue (#10517)

examples/text-classification/run_glue.py

@@ -406,12 +406,15 @@ def main():
 
     # Training
     if training_args.do_train:
+        checkpoint = None
         if last_checkpoint is not None:
             checkpoint = last_checkpoint
         elif os.path.isdir(model_args.model_name_or_path):
+            # Check the config from that potential checkpoint has the right number of labels before using it as a
+            # checkpoint.
+            if AutoConfig.from_pretrained(model_args.model_name_or_path).num_labels == num_labels:
                 checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
+
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         metrics = train_result.metrics