Implementation of Diakogiannis et al.'s ResUNet-a deep learning framework in Tensorflow 2.0 - originally written in MXNet
If you're training this yourself look at increasing batch size; I'm having to train locally on a P4000, which means a batch size of 2 with 256x256 images. Such a low batch size is known to have implications on accuracy, even with "corrective" regularisation elsewhere...
Tversky loss utilised instead of Tanimoto. Code for both functions provided.