This repo contains example code to train a ResNet18 or ResNet50 classifier on the Chaksu dataset. The classifier is implemented in pytorch-lightning. Furthermore, a script is provided for training on the Tübingen ML Cloud infrastructure.
You can train the classifier with default settings using
python train.py --experiment_name="chaksu"
If you want to run the code on the Tübingen ML Cloud, use the following command
sbatch --partition=gpu-2080ti deploy.sh
Start a tensorboard instance in the runs directory, and open tensorboard in your browser.
tensorboard --logdir='./runs'
If you are using the ML Cloud, start a tensorboard in a tmux shell using a specific port, e.g.
tensorboard --logdir=runs --port=2326
and then SSH onto the login node using port forwarding, i.e.
ssh -L 2326:localhost:2326 slurm
This will make tensorboard available on your local browser on localhost:2326.
Once checkpoints are written you can start testing the model using this command
python test.py --checkpoints_dir=<path-to-your-checkpoint-folder> --checkpoint_identifier='auc'
where --checkpoints_dir points to the actual experiment name and --checkpoint_identifier allows you to choose between the model with the best validation auc or the lowest validation loss. If the argument is omitted, the latest model is used by default. Selecting by AUC provides better results.