@dxo/vision

Developer preview — API unstable. Not a closed vision-classify gate.

Composable vision surface: ResNet (backbone / features), LinearHead / Classifier / compose (logits), defineLabelSpace / decodeClassification (labels).

ResNet-18 uses DXO-native state keys (stem.*, stage{n}.block{i}.*). Preview forward supports 32×32 NCHW → [N,512] features.

pnpm add @dxo/vision

import { ResNet, LinearHead, compose } from '@dxo/vision';

const backbone = new ResNet({ depth: 18 });
const model = compose(backbone, new LinearHead({ output: 10 }));

Pretrained weight assets and torch/pth → DXO safetensors conversion live in the separate @dxo/resnet package (scripts/). Runtime loadWeights only accepts DXO-key safetensors (same keys as new ResNet().state()).