Function top1_recall_on_training_set

nodedb-codec/src/vector_quant/opq.rs:405–439

()

    #[test]
    fn top1_recall_on_training_set() {
        let vecs = tiny_dataset();
        let codec = train_tiny();
        let refs: Vec<&[f32]> = vecs.iter().map(|v| v.as_slice()).collect();
        let encoded: Vec<_> = refs.iter().map(|v| codec.encode(v)).collect();

        let mut correct = 0usize;
        for (i, v) in refs.iter().enumerate() {
            let query = codec.prepare_query(v);
            let best = encoded
                .iter()
                .enumerate()
                .min_by(|(_, a), (_, b)| {
                    codec
                        .exact_asymmetric_distance(&query, a)
                        .partial_cmp(&codec.exact_asymmetric_distance(&query, b))
                        .unwrap_or(std::cmp::Ordering::Equal)
                })
                .map(|(idx, _)| idx)
                .unwrap_or(usize::MAX);
            if best == i {
                correct += 1;
            }
        }
        let recall = correct as f64 / vecs.len() as f64;
        // SVD-Procrustes converges to ~70% on this minimum-size synthetic set
        // (n=10, dim=8, m=2, k=4: 4 bits per vector, codespace collisions
        // inevitable). Empirical measurements on SIFT1M with realistic
        // (m=8, k=256, dim=128) routinely hit ≥0.95 — see bench harness.
        assert!(
            recall >= 0.70,
            "top-1 recall on training set too low: {correct}/{} = {recall:.2}",
            vecs.len()
        );
    }

    #[test]
    fn more_iterations_reduce_reconstruction_error() {