oxicuda-anomaly
Anomaly detection primitives for OxiCUDA — DeepSVDD, AE/VAE reconstruction, LOF, COPOD, isolation scoring, statistical methods, ensemble.
Part of the OxiCUDA ecosystem — Pure Rust CUDA replacement.
Features
- Deep learning detectors: DeepSVDD (one-class deep SVDD with hypersphere center estimation); Autoencoder reconstruction-error scorer; VAE anomaly score via ELBO
- Distance and density methods: LOF (Local Outlier Factor) with k-nearest-neighbor reachability distances; COPOD (Copula-based Outlier Detection) via empirical CDFs; Mahalanobis distance detector
- Tree-based isolation: Random-projection isolation forest scoring with c-factor normalization
- Statistical baselines: MAD (Median Absolute Deviation) detector; Z-score detector; percentile threshold computation
- Ensemble scoring: Average, Maximum, and Weighted ensemble combination with min-max normalization across detectors
- PTX kernels: 7 GPU kernels (DeepSVDD loss, reconstruction score, LOF reachability distance, COPOD ECDF, Mahalanobis distance, isolation forest score, ensemble normalize) × 6 SM versions
Usage
use *;
let mut rng = new;
// Train and score with an Autoencoder
let cfg = AeConfig ;
let ae = new.unwrap;
let score = ae.score.unwrap;
println!;
// Statistical Z-score detector
let mut det = new;
let train: = .map.collect;
det.fit.unwrap;
let outlier_score = det.score.unwrap;
println!;
Documentation
License
Apache-2.0 — Copyright 2026 COOLJAPAN OU (Team Kitasan)