Training is through train.py and inference is through predict.py.

This code will generate pseudo-labels on the fly which will be used for training multi-scale architectures.