ajdnik / imghash / 27193971176

88%

master: 90%

push

github

feat: DINOHash (#64)

* feat: add DINOHash perceptual hash with separate weights module

Adds DINOHash, a 96-bit perceptual hash backed by a frozen DINOv2 ViT-S/14
(with register tokens) backbone fused with a PCA-aligned linear head.
Inference runs in pure Go on top of gonum BLAS — no CGO, no ONNX runtime.
Hashes match the proteus-photos compute_hashes_onnx.py reference bit-exact
on 5/6 fixtures; the lena outlier differs by 1 bit due to PIL vs Go
bilinear sub-LSB pixel drift.

Model weights ship in a new sibling module (dinoweights) so importers of
imghash do not pay the ~85 MB embed cost unless they opt in. The
dinoweights module has zero requires beyond the stdlib; production code
exports a single Blob []byte variable consumed via WithSafetensorsBlob.
WithDINOWeights is provided for custom WeightsProvider implementations
(disk, network, alternate format).

ViT inference internals live in internal/vit (layer norm, attention, MLP,
patch embed, model forward). Preprocessing, safetensors parsing, bicubic
positional-embedding interpolation, and weight-map -> model loading live
in internal/dinov2. The root package exposes only the public surface:
DINOHash type, NewDINOHash, Calculate, Compare, Tensor alias,
WeightsProvider, ParseSafetensors, WithSafetensorsBlob, WithDINOWeights,
ErrNoWeights.

* docs: rewrite wiki

* docs: update wiki with dinohash

* docs: fix wiki formating

* fix: resolve lint issues

415 of 509 new or added lines in 14 files covered. (81.53%)

3646 of 3975 relevant lines covered (91.72%)

6751041.68 hits per line