A working+verified release

In this release, the training code was thoroughly evaluated. The training script is usage now. You need to set
the correct data path in the script. A bunch of other parameters can also be set. For convenience, docker images are also provided.

Demo scripts are provided for adapting this repo to your own application. These files are demo_*.py.

I spent some time getting netvlad to work on Nvidia's TX2 with TensorRT. A Utility is provided to turn the keras model (HDF5) to UFF which can be loaded on TensorRT for speeded up descriptor computation.