Published by WaifuFinder. manifest.json pins 5 model files by SHA-256;
the app verifies every one of them, plus the licence, attribution and validation files,
before installation.
WD SwinV2 Tagger v3 — Core ML
Native fixed-batch tagger for WaifuFinder, converted from SmilingWolf/wd-swinv2-tagger-v3.
WDTagger.mlpackage accepts RGB float32 [8,3,448,448] values in 0–255 and returns sigmoid probabilities [8,10861]. Use CPU and GPU compute units. All 24 SwinV2 attention blocks use native SDPA; cosine normalization, learned temperature, relative bias and shifted-window masks are preserved. Window and head axes share one head axis so all SDPA queries/masks are rank four. BGR permutation and normalization are fused; do not apply either in the caller. Resize longest side to at most 1024, center on a white square and resize bicubically to 448 before inference. Duplicate the last frame to pad tails.
wd-tagger-labels.csv is in exact output order: 4 rating, 8,106 general, 2,751 character heads. WaifuFinder emits only general tags at threshold 0.35 and the winning rating. validation.json records original PyTorch parity. manifest.json pins every model member and labels by SHA-256. The app compiles the package locally at installation.
Apache-2.0. See LICENSE and NOTICE.
Fused attention qualification
Release version: swinv2-sdpa-rank4-v1. All 24 SDPA query/key/value/mask inputs are checked to be rank four before staging. The math preserves the original cosine normalization, per-head temperature, position bias and shift masks.
Offline synthetic tests retain all 15 reference top heads for all 24 diverse frames. Maximum error over that extended set is 0.01569 (the prior FP16 baseline reaches 0.02138); both have one threshold-classification difference. Random-batch validation is recorded in validation.json. This is not a natural-image quality benchmark.
The initial rank-five trial encountered an internal MPSGraph memory-encoding assertion. This release removes that SDPA view/broadcast layout. Three model instances completed 72 concurrent mixed batches, and the exact native test sequence passed 15 repeated replays. A permanent two-model concurrency regression also passes. No exception is caught or suppressed, and no CPU fallback is used. The original runtime assertion could not be reproduced reliably, so its exact internal trigger remains unconfirmed.
- Downloads last month
- -