Validation model can be different from fit model in estimator class

[liuzh47]

Description

In current estimator class implementation, validation model is the same network as the training model self.net. Code of evaluate_batch() is shown below:

        data, label = self._get_data_and_label(val_batch, self.context, batch_axis)
        pred = [self.net(x) for x in data]
        loss = [self.evaluation_loss(y_hat, y) for y_hat, y in zip(pred, label)]

This assumption does not hold true in the general case. It is common to have a different validation model than the training model on many tasks, e.g., the language model.

   model_eval = nlp.model.AWDRNN(args.model, len(vocab), args.emsize, args.nhid, args.nlayers,
                                  args.tied, args.dropout, args.weight_dropout,
                                  args.dropout_h, args.dropout_i, args.dropout_e)
   model = nlp.model.train.AWDRNN(args.model, len(vocab), args.emsize, args.nhid, args.nlayers,
                                   args.tied, args.dropout, args.weight_dropout,
                                   args.dropout_h, args.dropout_i, args.dropout_e)

Please refer to (https://github.com/dmlc/gluon-nlp/blob/master/scripts/language_model/word_language_model.py#L154) for a detailed reference.