Changes
- release: v0.6.0 — TensorRT and OpenVINO INT8 export, ByteTrack tracking @Gabriellgpc (#84)
- release: latency-drift sessions, and drop the Sponsoring section for now @Gabriellgpc (#80)
- release: pause the sponsor funnel, and namespace the release labels @Gabriellgpc (#77)
New Features
- feat(train): train on pseudo-labels alongside real COCO @Gabriellgpc (#73)
- feat(analysis): measure how far a dataset is from COCO @Gabriellgpc (#68)
- feat(pretrain): self-supervised pretraining on unlabelled images, with DenseCL @Gabriellgpc (#67)
- feat(train): initialise the backbone from a distilled checkpoint @Gabriellgpc (#66)
- feat(distill): log to TensorBoard or W&B, and report the cosine directly @Gabriellgpc (#62)
- feat(export): onnx, openvino and tensorrt at any input size, measured @Gabriellgpc (#51)
- feat(training): distil DINOv3 into the backbone from cached features @Gabriellgpc (#50)
- feat(tools): pseudo-label COCO with an open-licensed teacher @Gabriellgpc (#48)
- feat(tracking): deep hm-sort as the built-in tracker @Gabriellgpc (#49)
- feat(export): a self-contained onnx graph for per-object embeddings @Gabriellgpc (#46)
- feat(export): onnx export for embeddings, and for detections beside them @Gabriellgpc (#44)
- feat(embed): vision feature embeddings, and one pass for detections + embeddings @Gabriellgpc (#41)
- feat(api): name the COCO class ids with a COCOClass enum @Gabriellgpc (#39)
Bug Fixes
- fix(distill): validate on unaugmented data, so the metric measures the model @Gabriellgpc (#64)
- fix(distill): shards do not hold shard_size images, and the reader assumed they do @Gabriellgpc (#60)
- fix(data): drop zero-area COCO annotations before they reach a transform @Gabriellgpc (#56)
- fix(data): ignore COCO crowd regions instead of teaching them as background @Gabriellgpc (#53)
- fix(recipes): the COCO learning rate was never validated @Gabriellgpc (#55)
- fix(train): make
yolonas train able to run the COCO recipe @Gabriellgpc (#52)
Documentation
- docs: remove the Sponsoring section until GitHub Sponsors is set up @Gabriellgpc (#79)
- docs(benchmarks): three YOLO-NAS-S latency sessions, to measure drift @Gabriellgpc (#78)
- docs(sponsors): take the funding links down until payouts are fixed @Gabriellgpc (#76)
- docs(plan): stage 2 done -- the plumbing gate, and unlabeled2017 pseudo-labelled @Gabriellgpc (#72)
- docs(plan): survey the pretraining, distillation and fine-tuning literature @Gabriellgpc (#65)
- docs(plan): keep the feature cache, and fix the trigger for deleting it @Gabriellgpc (#71)
- docs(plan): the stage 1 gate cleared, and the margin was still growing @Gabriellgpc (#70)
- docs(plan): self-supervised pretraining, and how to decide whether it is worth it @Gabriellgpc (#69)
- docs(sponsors): point funding at the CondadosAI org and add SPONSORS.md @Gabriellgpc (#54)
- docs(plan): reclaim the 279 GB feature cache once stage 1 has passed its gate @Gabriellgpc (#63)
- docs(plan): record the alternatives evaluated, and why each was rejected @Gabriellgpc (#61)
- docs(plan): the feature cache is disk-bound, and the plan assumed writing was free @Gabriellgpc (#59)
- docs(plan): say what the calibration gate does not cover @Gabriellgpc (#57)
- docs: plan the Apache-2.0 COCO weights, and measure the teacher cost that shapes it @Gabriellgpc (#47)
- docs(roadmap): replace the dataloader guess with measurements @Gabriellgpc (#40)
- docs: read the license badge from the repo, not from PyPI @Gabriellgpc (#34)
- docs(funding): add sponsorship options @Gabriellgpc (#33)
Maintenance
- ci(release): namespace the version labels so Dependabot cannot set them @Gabriellgpc (#74)
- chore: roll up dev into main for the next release @Gabriellgpc (#38)
- test(callbacks): pin that CloseMosaicCallback reaches the dataloader workers @Gabriellgpc (#58)
- perf(data): mixup blend, the benchmark harness, and what it found on the training machine @Gabriellgpc (#45)
- perf(data): fuse crop, flip and affine into a single warp @Gabriellgpc (#43)
- perf(train): cut the dataloader's two wasteful stages, and add channels_last @Gabriellgpc (#42)
- chore: Bump the python-packages group with 23 updates @dependabot[bot] (#35)
- chore(spaces): move the Gradio demo to modern-yolonas 0.5.0 @Gabriellgpc (#37)
- chore(lint): pin the ruff rule set so an upgrade cannot widen it @Gabriellgpc (#36)
Contributors
@Gabriellgpc, @dependabot[bot] and @claude