Crate oxicuda_ssl 

oxicuda-ssl — Self-supervised learning primitives for OxiCUDA.

Pure-Rust implementation of the four canonical SSL families, suitable for CPU simulation and PTX kernel generation for GPU execution.

Architecture

oxicuda-ssl
├── contrastive/      — SimCLR (NT-Xent), MoCo (memory-bank InfoNCE)
├── non_contrastive/  — BYOL (cosine), Barlow Twins, VICReg
├── masked/           — MAE (random patch mask + reconstruction MSE)
├── clustering/       — SwAV (Sinkhorn-Knopp), DINO (centred + sharpened CE)
├── augment/          — Color jitter, multi-crop helpers
├── metrics/          — Uniformity, alignment, effective rank, collapse score
├── momentum/         — EmaUpdater for momentum-encoder schemes
├── head/             — MlpProjector, PredictorHead
├── error             — SslError / SslResult
├── handle            — SslHandle (SmVersion + LcgRng)
└── ptx_kernels       — GPU PTX kernel strings

Modules

augment

Standard SSL data augmentation helpers operating on [C, H, W] CHW tensors.

clustering

Clustering SSL losses: SwAV (Sinkhorn-Knopp normalised assignments), DINO (centred + sharpened student-teacher cross-entropy), iBOT (masked image modeling with online tokenizer), and DeepCluster / DeeperCluster (k-means pseudo-label clustering).

contrastive

Contrastive SSL losses: SimCLR (NT-Xent) and MoCo (memory-bank InfoNCE).

error

Error types for oxicuda-ssl.

handle

Session handle for oxicuda-ssl.

head

Projection and predictor MLP heads used by SSL pipelines.

masked

Masked SSL losses: Masked Autoencoder (MAE) random-patch dropping + reconstruction MSE; SimMIM L1/L2 reconstruction with block & random masking; data2vec joint-embedding masked prediction (Baevski et al. 2022); BEiT discrete-token prediction (Bao et al. 2021).

metrics

Representation-quality metrics for self-supervised learning.

momentum

Momentum-encoder utilities (EMA target network update + scheduling).

non_contrastive

Non-contrastive SSL losses: BYOL (cosine target), Barlow Twins (cross-correlation), VICReg (variance + invariance + covariance), SimSiam (stop-gradient cosine), and MSN (masked siamese networks).

prelude

Convenience re-exports for common SSL types.

ptx_kernels

PTX GPU kernel sources for self-supervised learning operations.

ssl

High-level SSL model structs that own their own weights.