Booleanization

Tsetlin machines consume Boolean features. Turning raw data into Booleans is the single most important modelling decision, because the literals are the vocabulary of the learned rules. torchtsetlin.data provides encoders that follow the scikit-learn fit / transform protocol and are nn.Modules (they move to the GPU and are saved in state_dict()).

Continuous features: thermometer encoding

ThermometerEncoder(n_bits, strategy) turns each value v into bits [v >= t_1, …, v >= t_n] for increasing thresholds. Two literals then express a range: age>=30 AND NOT age>=50.

Thermometer encoding: each bit answers “is the value at least this much?”, so the 1s fill up from the left as the value grows. Any range is then just two literals.

enc = tt.data.ThermometerEncoder(n_bits=8, strategy="quantile").fit(x_train_float)
xb = enc(x_train_float)                       # (N, F * 8) bool
model.feature_names = enc.feature_names(column_names)   # readable rules

Strategies: "quantile" (equal-frequency, robust default), "uniform" (equal-width, pass value_range=(0, 255) for images), "unique" (evenly spaced ranks of the unique training values, like tmu's StandardBinarizer).

Categorical features

OneHotEncoder produces one Boolean per category (menopause=ge40), as in the book's breast cancer example. BitPlaneEncoder gives the binary digits of integers when you want compact codes (even numbers are NOT bit0).

Images

• Binarizer(0.3) — fixed threshold (the original MNIST recipe: pixel > 0.3).
• AdaptiveThresholdEncoder(block_size=11, C=2, method="gaussian") — local thresholding (OpenCV's ADAPTIVE_THRESH_GAUSSIAN_C), used for MNIST/Fashion-MNIST in the convolutional TM papers.
• ColorThermometerEncoder(n_bits=8) — per-channel thermometer, (N, 3, H, W) → (N, 24, H, W), the standard CIFAR encoding for TM composites.
• BitPlaneEncoder(8) — 8 place-value planes per channel.

load_mnist_boolean / load_torchvision_boolean apply any of these to a torchvision dataset and return Boolean tensors ready for the GPU.

Text and sets

Bag-of-words presence bits are the classical text encoding. HypervectorEncoder maps token ids to sparse random hypervectors and bundles them with OR (optionally binding word position with a cyclic shift), a compact alternative for large vocabularies.

Composition

pipeline = tt.data.Compose(tt.data.ThermometerEncoder(4), tt.data.Flatten())
xb = pipeline.fit_transform(x)
pipeline.output_size(n_columns)

Decoding rules

ThermometerEncoder.decode_range and ConvTsetlinMachine.clause_region translate included bits back to intervals, and interpret.aggregate_literal_importance sums literal importances over the bits of one original feature.