Repository navigation
refactor: move YOLO out of core, keep the reusable detection pieces - #1437
Merged
Merged
Conversation
The tape records TensorOps.softmax(t, dim) with parameters["dim"], but the StableHLO converter read parameters["axis"] and fell back to the last axis, so every exported softmax over a non-last dimension reduced over the wrong one (found exporting a YOLOv8 DFL head: softmax over 16 bins ran over the 8400 anchors). logSoftmax, recorded as softmax with log = true, was exported as a plain softmax; it now lowers to x - max - log(sum(exp(x - max))). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The YOLOv8 graph, config and decoder leave core for their cartridge blueprint (detection-yolov8-iree), which owns the model end to end. - skainet-data-media: Letterbox (fit / fillNchw / toSource) and BoundingBox, ScoredBox, nonMaxSuppression, with tests; the module gets the publish plugin (it was never on Maven Central). - skainet-model-yolo removed with its wiring (settings, dokka, tensor-tools). - ModelArchitecture.YOLO deprecated; yolo.* GGUF key aliases dropped. - YOLO wording out of comments, tests (MultiScaleHeadGraphDslTest) and docs; CHANGELOG [Unreleased]. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
📖 Documentation Preview The documentation has been built successfully for this PR. Generated Files:
Artifacts:
This comment will be updated automatically when the PR is updated. |
|
📖 Documentation Preview The documentation has been built successfully for this PR. Generated Files:
Artifacts:
This comment will be updated automatically when the PR is updated. |
michalharakal
added a commit
that referenced
this pull request
Oct 21, 2026
CHANGELOG [0.62.1] - 2026-10-21 over everything develop holds since 0.62.0: the FFM Q8_0/Q4_0/Q5_0/Q5_1 binding fix (#1436) as the headline, the YOLO move-out (#1437), the StableHLO softmax axis and logSoftmax fixes, the compile-flexibility explanation (#1432) and the SKEEP-008 draft (#1434). VERSION_NAME, CITATION, Antora version, README bom coordinate, What's New and contributors, Android tutorial coordinates; regenerated operator and kernel-support matrices.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Model-specific detector code leaves core; the reusable detection pieces stay. Two commits, rebased onto
develop (0.62.0-dev), touched-module
jvmTest+ repo-wideapiCheckgreen.What moves out
skainet-model-yolo(graph, config, decoder, preprocess) is removed; it lives on in its cartridgeblueprint —
detection-yolov8-ireein SKaiNET-cartridge-blueprints — which owns the model end to end(ADR-015). The 0.58.0–0.60.0 artifacts stay on Maven Central.
ModelArchitecture.YOLOis deprecated,the
yolo.*GGUF keys are no longer read, tests/docs reworded.What stays, generalized (
skainet-data-media, now published)Letterbox: aspect-preserving fit into a square canvas without a platform image —fit(width, height, targetSize),fillNchw(rgb888, padValue = 114)(bilinear,[0,1], NCHW),toSource(box)back-mapping, clamped.BoundingBox/ScoredBox/nonMaxSuppression: pixel boxes withiouandfromCenter, greedyclass-aware NMS with
maxDetections.Fixes riding along (
skainet-compile-hlo)softmaxlowering honouredparameters["axis"]but the tape recordsdim, so every exportedsoftmax(t, dim)reduced over the last axis. Readsdimnow (axisstill accepted).logSoftmax(recorded assoftmax(log = true)) exported as a plain softmax; now lowers tox - max - log(sum(exp(x - max))).Verification
apiCheckgreen after the rebase.host (eager and IREE llvm-cpu, head rtol/atol 1e-2, identical post-NMS detections) and passes the
golden run on a Pixel 8a (CPU and Vulkan packs, identical scores).
🤖 Generated with Claude Code