A sample implementation of deeploco.
train_script.py trains a neural net to do localization using simulated data generated from a z-stack from the 2016 SMLM challenge.
You'll need a machine with a (reasonably) powerful GPU to train quickly (set use_cuda=True).
To try this on a new dataset you'll need a z-stack (see empirical_sim.py) and to make sure that the simulated data looks as similar as possible to the real data. This could be quite difficult: you'll need to adjust many hardcoded values in empirical_sim.py, as well as the generative model settings (in train_script.py). You might also need to modify the learning rate schedule of the network.
localize.py gives an example of how to use a pretrained network to do localization.