checkpoint_model ignores best model if just one metric stays the same

[Xwaler]

This issue concerns the checkpoint_model parameter in every trainable component (DietClassifier, ResponseSelector, RulePolicy, etc...)
In order to save a model that has improved in every metric except one that has staled (can even stale at 1.00 accuracy !), the _does_model_improve function could compare with a superior or equal sign rather than superior strict.

rasa/rasa/utils/tensorflow/models.py

Line 483 in 0ccfa20

float(current_results[key]) > self.best_metrics_so_far[key]

all_improved = all(
            [
                float(current_results[key]) >= self.best_metrics_so_far[key]
                for key in self.best_metrics_so_far.keys()
            ]
        )

[sara-tagger]

Thanks for the issue, @alopez will get back to you about it soon!

You may find help in the docs and the forum, too 🤗

[tttthomasssss]

@dakshvar22 assigned as reviewer.