Notebooks¶
End-to-end walkthroughs as Jupyter notebooks. The pages below show the code and the explanations; the notebooks are stored without outputs, so run one to see its results. The easiest way is Google Colab: open a notebook with its badge, and the first cell installs torchtsetlin from PyPI. The datasets download on first use.
| Notebook | Open | Data | Model | What it is about | Runtime |
|---|---|---|---|---|---|
| Iris, a readable model |  |
Iris (scikit-learn) | TsetlinMachine |
Thermometer encoding, and reading the learned rules back as intervals in centimetres | ~1.5 min (CPU) |
| MNIST on the GPU | |
MNIST (torchvision) | TsetlinMachine, ConvTsetlinMachine |
Convolutional clauses, GPU throughput, and the batch-size / feedback-fidelity trade-off | ~1 min |
| CIFAR-10 | |
CIFAR-10 (torchvision) | ConvTsetlinMachine |
Booleanizing colour photographs, and where the accuracy ceiling is actually set | ~2 min |
| Convolutional regression (C-RTM) | |
synthetic (the paper's Fig. 3) | ConvRegressionTsetlinMachine |
C-RTM (ICMLT 2021): how a TM outputs a number, and what its clauses say afterwards | ~6 min |
| CTM-UNet segmentation | |
synthetic CamVid-like scenes | TsetlinMachine (dense per-pixel head) |
CTM-UNet (ISTM 2025): which of its ingredients carry the result | ~1 min |
| Graph Tsetlin machine | |
synthetic sequences, MNIST | GraphTsetlinMachine |
Granmo et al. (2025): deep clauses over graphs, reading them back, and depth 1 = convolution | ~1 min |
| Semantic segmentation | |
synthetic street scenes | SegmentationTsetlinMachine |
Per-pixel labels: the row-prior floor, context from position literals and a pyramid, explaining a pixel, patches_per_commit |
~3 min |
| Polyp segmentation (Kvasir-SEG) | |
Kvasir-SEG (downloaded) | SegmentationTsetlinMachine |
Real colonoscopy images: a location-prior floor, class-balanced feedback, early stopping, and clauses that expose the model's shortcut | ~6 min |
Running in Colab¶
- For notebooks 2–8 choose Runtime → Change runtime type → T4 GPU before running. A T4 is slower than the RTX 3090 and 3080 the runtimes above were measured on, so expect them to take longer. Notebook 1 runs on the CPU.
- The Colab links open the notebooks from the
mainbranch on GitHub. - Colab already has torch, torchvision, scikit-learn, matplotlib and pandas; the setup cell only adds torchtsetlin.
Running locally¶
pip install torchtsetlin jupyterlab # or pip install -e ".[dev]" in a checkout
jupyter lab examples/notebooks
MNIST, CIFAR-10 and Kvasir-SEG download to data/ at the repository root on first run. Every notebook also
runs on the CPU, but notebooks 2–8 are written for a GPU and are much slower without one.
The source files are in
examples/notebooks/.