Currently if a dict is passed to the hyperparameters param in TrainingStep constructor, this dict replaces the estimator one.
This way, not only all hyperparameters passed to the estimator constructor are removed, but the reserved ones (sagemaker_* ) are too.
So currently, in order to use hyperparameters in TrainingStep, you have to figure out which reserved hyperparams are required for the job to work and provide them explicitely.
It would be useful for the hyperparams passed to TrainingStep to update, rather than replace, the estimator hyperpars' dict.