example_scripts

Example scripts

Short, complete programs, one feature of LION each. For a step-by-step introduction to the library, see the demos (demos/).

Run them from the repository root, in order (some use the models of earlier ones), e.g. python scripts/example_scripts/01_supervised_training.py. Each one has a # %% cell per step, so an editor that knows cells (VS Code, Spyder, Jupyter) runs them a step at a time.

They need:

• a GPU: LION's CT operators project with tomosipo and ASTRA's CUDA code;
• the processed LIDC-IDRI data in LION_DATA_PATH (see Datasets);
• LION_RESULTS_PATH: they write to $LION_RESULTS_PATH/examples/.

By default they train a few epochs on a few hundred slices, to show they work. Change the settings at the top for real training (more epochs, all the data).

Script	What it shows
01_supervised_training.py	Supervised training of FBPConvNet: an experiment, a model, an optimizer and a learning-rate scheduler, validation (keeping the best model), checkpoints, saving. Any supervised model is one line away.
02_learned_primal_dual.py	An unrolled method, LPD: a model that takes the measurements and trains through the CT operator; changing a model's parameters.
03_self_supervised_noise2inverse.py	Self-supervised training with Noise2Inverse (a U-Net), without clean images; the other self-supervised solvers are used the same way.
04_pnp_denoiser.py	Plug-and-Play: a DRUNet trained as a Gaussian denoiser, then the prior of a PnP-ADMM reconstruction, compared with FDK.
05_evaluate_and_compare.py	Evaluating methods on an experiment's test data, and a comparison (tables, scores, figures) of the models of 01, 02 and 04 with SIRT, against FDK.
06_resume_and_fine_tune.py	A stopped training resumed from its checkpoint; a trained model fine-tuned on another problem, its record naming its original.
07_publish_and_reproduce.py	Publishing a trained model with its records, then making it again, and training it again with a change, from the package alone.
08_your_own_pieces.py	Your own torch model (TorchModel), forward operator (a blur) and dataset, trained with LION's solver.
photocurrent_mapping.py	Photocurrent mapping, a modality other than CT: compressed-sensing baselines (FISTA, SPGL1) and PnP-ADMM on subsampled measurements. Needs the PCM data in data_dir, at the top of the file.