After debugging an issue with keras with @jdu, Jeff suggested we need to redesign some parts of how the tests are run.

• We shouldnt be using a virtualenv with tox where you pre-install a requirements.txt. tox creates a virtualenv from the venv you instantiate in the repo. This means that any site-packages are copied into the new venv that tox creates for testing purposes. Basically tox tears down its virtualenvs it creates on each run, but if the venv is based on another venv, it will copy site-packages installed from the parent venv, so when you run tox this way you’re not actually doing a fresh clean install of the package.
• We need to fix that Makefile so that you DO NOT overwrite existing referenced versions of the deep_reference_parser wheel/pkg so that changes to the package don’t break existing processes depending on it.