Ensure training each backbone and head attains good accuracy

[ravinkohli]

Currently, in the test for networks (here) combinations of backbone and head do not achieve 80% accuracy with the default configuration. We would like to fix the search space with good defaults so as all the combinations achieve at least 80%.
Originally posted by @franchuterivera in #78 (comment)

[franchuterivera]

This has been addressed by #82