The easiest way to quickly setup the repository is with using Docker. In the terminal, move to the docker folder and run:
bash docker_build.sh
to setup the docker environment.
Download the data at??? You'll need to at the location of your data to the files docker_run.sh and docker_run_jupyter.sh. In both files, change the line:
-v path/to/your/data:/data \
to wherever you've downloaded and unzipped the data.
You can run the training within a docker container. First, run:
bash docker_run.sh
To open a terminal within the container. Second, start the training with
python exe_train_models.py
This will start training for the ResNet-18. If you want to train another model, change the strings in the USED_MODELS list. If your machine has more than one GPU, update the list AVAILABLE_GPUS. Several experiments will be run in parallel.
During training, a folder is created in experiments. It contains the model.pt file, as well as the training log (seg_log_model.json), the passed options for the model (seg_opt.json), and a list of which images were used for training, and which for validation (split.json).