loads the tokenizer for each checkpoint, to solve the reproducability issue

examples/run_glue.py

@@ -463,6 +463,7 @@ def main():
         for checkpoint in checkpoints:
             global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
             model = model_class.from_pretrained(checkpoint)
+            tokenizer = tokenizer_class.from_pretrained(checkpoint)
             model.to(args.device)
             result = evaluate(args, model, tokenizer, prefix=global_step)
             result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())