Out-of-the-box inference without training?

Is it possible to instantiate `DeepBiosphere` with your trained weights and perform inference without needing to train the model again ourselves? 

This is related to #4 somewhat (in the sense that an `.onnx` or `.pt` could be obtained as one of those potential inputs) but thought I'd keep it separate to avoid any confusion. 