Don't add tf summary when calling AddEvalMetric.

Eval metric summaries are added inside trainer.py.

As a result, eval metrics are no longer reported by 'control' job, but that actually removes some clutter instead because control should be identical to train.

PiperOrigin-RevId: 213532190

lingvo/core/base_model.py

@@ -599,7 +599,6 @@ def AddEvalMetric(self, name, value, weight):
     """
     if name in self._eval_metrics:
       raise ValueError('Metric %s has already been defined.' % name)
-    summary_utils.scalar(self.params, name, value)
     self._eval_metrics[name] = (value, weight)
 
   @property