Tutorials¶

Five end-to-end examples live in the tutorials/ folder of the repository. Each is a script with its configuration, from a four-feature classifier that trains in seconds to a cross-validated benchmark with fine-tuning that runs for hours.

Tutorial	Task	Families	Shows	Run time
Iris	`multi-class`, 3 classes, 150 rows	`linear_cov`	a whole run in seconds; a shared projection as the class map	about 11 s on a CPU
California housing	`regressor`, 20,640 rows	`legendre` or `rbf`	feature engineering, an output head and the end-to-end pass, a comparison with gradient boosting	about 45 s on a GPU
Concrete	`regressor`, 1,030 rows	`linear_cov`, `quadratic`, `cubic`, `polyquad`	four families side by side, a 30-layer network on a small dataset	about 8 min on a CPU
CCPP	`regressor`, 9,568 rows, 5×2 cross-validation	power-basis and `legendre`	the published benchmark protocol, the squash, the three fine-tuning passes	12 min on a CPU to 14.7 h on a GPU, by config
Otto	`multi-class`, 9 classes, 61,878 rows	`polyquad`	class weights, a projection per candidate, the per-layer fine-tune, a head over 93 survivors	about 2 min on a GPU

The run times are single runs of the shipped configurations; the CCPP range covers its configs, each over all ten folds.

Running them¶

The tutorials are part of the repository, not of the installed package. Clone the repository and install it with the test extra, which adds pandas and the viz extra (see Installation):

git clone https://github.com/kvoyager/torchsonn.git
cd torchsonn
pip install -e ".[test]"

Then run a tutorial from the repository root:

python -m tutorials.iris.iris_recognition
python -m tutorials.california_housing.california_housing --config-name california_housing_legendre_finetune \
    train.early_stop_source=train train.checkpoint_dir=checkpoints/california
python -m tutorials.concrete.concrete train.checkpoint_dir=checkpoints/concrete
python -m tutorials.ccpp.ccpp
python tutorials/otto/otto_classification.py

Every training run writes its checkpoints and train.log to a folder of its own, named after its start time, inside the configuration's train.checkpoint_dir; each script prints that folder at the end (see Training).

Iris loads its data from scikit-learn. The others download theirs on the first run and cache it: California housing through scikit-learn, the others into a data folder next to the script.

Every tutorial except Otto reads its configuration with Hydra, so any key can be overridden on the command line (see Configuration):

python -m tutorials.iris.iris_recognition train.max_layer_count=3
python -m tutorials.ccpp.ccpp --config-name ccpp_legendre train.device=cuda

Every tutorial except Otto draws its network with Graphviz, which needs the viz extra and the Graphviz programs. Iris, Concrete and CCPP draw into .svg files next to the script, overwriting the copies in the repository; California housing draws into its run folder. Each tutorial page explains its command's extra settings.

Checked against TorchSONN 0.1.5.