In this repository, we provide the code to train a segmentation model based on an unet architecture. There can be changed several parameters including the loss function in order to test the influence of the power term during training. It is necessary to unzip the data folder before running the notebook.
The folder "examples" provides the results of running "Demo" notebook with some loss functions: crossentropy, classical Jaccard and power Jaccard with p = 2.
To run the demo notebook, please install:
- Tensorflow 1.12.0
This code is provided as a complementary material of the paper "On power losses for semantic segmentation" by David Duque-Arias, Santiago Velasco-Forero, Jean-Emmanuel Deschaud, Francois Goulette, Andres Serna, Etienne Decenciere and Beatriz Marcotegui from PSL Research University and Terra3D Research.
For any additional request, do not hesitate to contact us a david.duque@mines-paristech.fr
Note: If Github rendering of Jupyter notebooks is not working, please go to https://nbviewer.jupyter.org/github/daduquea/powerLosses/blob/main/Demo.ipynb?flush_cache