Estimator developers should clearly been informed why their tests are failing. We weren't able to do this with assert_raises until we very recently backported from python 3 the ability to use it as a context manager and provide an assertion error message. These in particular should be improved.

@amueller please feel free to expand upon this.