subsume

Geometric box embeddings for modeling containment ("is-a") and entailment relationships.

Dual-licensed under MIT or Apache-2.0.

Quickstart

[dependencies]
subsume = "0.1.0"
subsume-ndarray = "0.1.0"
ndarray = "0.15"

use subsume_ndarray::NdarrayBox;
use ndarray::array;

// Box A: [0,0,0] to [1,1,1]
let premise = NdarrayBox::new(array![0., 0., 0.], array![1., 1., 1.], 1.0)?;

// Box B: [0.2,0.2,0.2] to [0.8,0.8,0.8] (inside A)
let hypothesis = NdarrayBox::new(array![0.2, 0.2, 0.2], array![0.8, 0.8, 0.8], 1.0)?;

// Probability that A contains B
let p = premise.containment_prob(&hypothesis, 1.0)?;
println!("P(B ⊆ A) = {:.2}", p);

Features

• Gumbel Box: Probabilistic boxes for training stability
• Backends: ndarray (CPU) and candle (GPU/Metal)
• Training: Volume regularization, temperature scheduling
• Inference: Fast containment and overlap scoring

See docs/ for mathematical foundations and research details.