This repository was created for the course "Computational Intelligence Lab" at ETH ZÃrich. It contains code used to predict where roads lie in given aerial images.
The code was written by the group 'Cillers' consisting of Mike Boss, Jonas Passweg and Jonathan Ehrat.
To run the code a running version of Python 3 is needed. The version we used is Python 3.7.4. Information about how to get Python on your system should be readily available on the internet.
Additionally you will need some Python packages. These can be installed via Pip, the package installer for Python.
The needed packages are found in the 'requirements.txt' file and can easily be installed with the command pip install -r requirements.txt
To have all the training data we used you can download additional data from the github repository https://github.com/ehratjon/cil_2021/tree/master.
To get access to our trained models, download them from the polybox https://polybox.ethz.ch/index.php/s/rDB7FL4XGeSpEp7 and add them to lightning_logs/<model_name>, where <model_name> is the name of the checkpoint (e.g. UNetSpatialDilated_data)
To run the code there are various ways.
Reproducing the model trained for the report simply run python Nested_Checkpoint.py to load the best model or run python cillers_best_training.py to train the model from scratch.
To interactively play with the provided models there is an interactive jupyter notebook.
Start a notebook server with the lightning.ipynb notebook using jupyter-notebook cillers_best_notebook.ipynb.
This opens the notebook in your browser.
If you wish to train and predict a specific model you can start a python script with python training_scripts/model.py, where model.py is the filename of the model you want to train.
If you would like to run and test all models on the Leonhard cluster, connect to the cluster and use python run.py to have each model submited as a job.