Crate oxicuda_vision 

oxicuda-vision — Vision Transformer & CLIP primitives for OxiCUDA.

Pure-Rust CPU reference implementation providing:

• patch_embed: strided Conv2D patch embedder, sinusoidal & learnable positional encodings.
• vit: ViT block (pre-norm MHSA + MLP), encoder stack, full ViT model, and the Swin Transformer windowed / shifted-window block.
• convnext: ConvNeXt modern-CNN block (depthwise conv + channel LayerNorm + inverted-bottleneck + layer scale).
• clip: CLIP vision encoder, projection head, InfoNCE contrastive loss.
• augment: geometric, photometric, and normalisation image augmentations, plus MixUp / CutMix batch mixing regularisers.
• imgproc: classical image processing — Sobel gradients and the Canny edge detector, binary/grayscale morphology, union-find connected-component labelling, and the Hough line transform.
• fpn: Feature Pyramid Network (lateral 1×1 convolutions + top-down pathway).
• detection: RoI Align, DETR decoder, bipartite set matching, IoU / GIoU / DIoU / CIoU box-regression losses, the RTMDet detector (CSPNeXt backbone + PAFPN neck + decoupled head + SimOTA-lite cost), and the OWL-ViT open-vocabulary detector (per-patch image-text matching).
• segmentation: the Segment Anything Model (SAM) — ViT image encoder, prompt encoder, and a two-way transformer mask decoder.
• ssl: the DINOv2 self-supervised distillation recipe — ViT backbone ([CLS] + patch tokens), weight-normalised prototype head, centred / sharpened teacher-student DINO loss, EMA teacher, centering buffer, and the iBOT masked-patch term.
• text: the CLIP Transformer text encoder — token + positional embeddings, causal self-attention blocks, EOS pooling, and joint-space projection.
• pointcloud: the Point Transformer vector self-attention layer over kNN neighbourhoods.
• losses: focal loss (sigmoid & softmax) and soft Dice segmentation loss.
• ptx_kernels: 7 GPU PTX kernel string generators (SM 7.5–12.0).

No CUDA SDK dependency; all forward passes run on CPU f32 tensors using flat row-major Vec<f32> layouts.

Re-exports

pub use error::VisionError;

pub use error::VisionResult;

pub use handle::LcgRng;

pub use handle::SmVersion;

pub use handle::VisionHandle;

Modules

augment

Image augmentation pipeline for CHW tensors.

blocks

EfficientNet-style building blocks.

clip

CLIP (Contrastive Language–Image Pre-Training) vision module.

convnext

ConvNeXt modern-CNN components.

detection

Object detection components.

error

Error types for oxicuda-vision.

fpn

Feature Pyramid Network (FPN) components.

handle

Session handle for oxicuda-vision.

imgproc

Classical image-processing primitives operating on flat f32 buffers.

losses

Loss functions for classification, detection, and segmentation heads.

optimize

Inference-time optimization passes for vision models.

patch_embed

Patch embedding for Vision Transformers.

pointcloud

Point-cloud neural network primitives.

prelude

ptx_kernels

PTX GPU kernel sources for vision model operations.

segmentation

Image segmentation models.

ssl

Self-supervised learning (SSL) methods for vision backbones.

text

Text encoders for vision-language models.

vit

Vision Transformer (ViT) components.