Question about loss in Gluon Estimator

[harupy]

In the current implementation, loss is monitored only when self._train_metrics is falsey (meaning train_metrics in Estimator.__init__ is unspecified). I'm just curious if this is an intended behavior or not.

https://github.com/apache/incubator-mxnet/blob/4405906977e99c3328b0879a6dbe4ce56bf30d57/python/mxnet/gluon/contrib/estimator/estimator.py#L240-L246

[harupy]

In mxnet 1.5.x, loss is always monitored regardless of self.train_metrics (line 186 ~ 188).

https://github.com/apache/incubator-mxnet/blob/ec36dfe3af277c55c81646302e892895cb5908ac/python/mxnet/gluon/contrib/estimator/estimator.py#L180-L190

[szha]

cc @roywei