Train the UNet with swift run TrainUNet. This saves weights in "checkpoints/".
Do inference using this Colab notebook (which imports DeepBeeTracking). It is configured to load some pretrained weights from the internet. You can edit it to point at your own weights.