Py2/Py3 Models
In order to run Dan's models in Python 2 (for ROS), I have to do some hacky workarounds. First I have to load the torch model in a python 3 torch 1.6, then save just the ['state_dict'] portion of the model. That's the only way a Python 2 Torch 1.4 will read the model. Then, for LightningModel, I have to convert the parameters back here.
Currently, the training would work without intervention if just the state_dict portion of the model was saved to the checkpoint.