Product roadmap
September 7, 2026 · View on GitHub
Updated: 2026-06-09 · Current release: v0.8.0
BNNR is a PyTorch vision toolkit focused on model diagnostics first (bnnr analyze), then saliency-guided augmentations (ICD/AICD), with optional training and detection adapters.
Start here: analyze.md · sample HTML report
Shipped (baseline)
| Area | What exists today |
|---|---|
| Analyze | bnnr analyze + portable HTML/JSON for classification and multilabel; sample report |
| Training | CLI train / demo / quickstart; presets; live dashboard |
| XAI aug | ICD, AICD, branch search; plugin_icd.md for custom loops |
| Detection | Train API + bbox augs; Ultralytics adapter; detection.md — no analyze for detection yet |
| Benchmarks | CIFAR-10 demo CNN (3 seeds): no aug vs RandAugment vs BNNR — benchmarks.md |
| Integrations | pytorch-grad-cam loop, Ultralytics quickstart — integrations.md |
| Torchvision analyze | Python example: torchvision_analyze_cifar10.py |
Q2 2026 (June–August) — credibility + analyze adoption
| Planned | Description |
|---|---|
| Analyze-first docs | README and getting_started.md: try analyze before train; githack sample links |
| Augmentation benchmark | ResNet18 / Imagewoof (fine-grained, low-data, from-scratch): 5 seeds; baselines RandAugment + TrivialAugment; reproducible script benchmarks/reproduce_imagewoof.sh (free-T4 friendly) |
| Torchvision → analyze | Prominent golden path: pretrained classifier → HTML report; runnable example in README |
| Repo discoverability | GitHub description, website link, topics (maintainer ops) |
| Contributor templates | Issue forms: benchmark proposal, “Who uses BNNR” showcase |
| Upstream docs | Official link in Ultralytics docs (after PR merge); grad-cam ecosystem follow-up |
| HuggingFace Space | Minimal Gradio demo (sample report or analyze workflow) |
Q3 2026 (September–November) — analysis depth + ecosystem
| Planned | Description |
|---|---|
| HuggingFace models | First transformers path for analyze (e.g. ViT) — documented example, not full Hub integration yet |
| Dataset insight | “What to label next” / outlier buckets for manual review |
| More XAI methods | Better stability + more methods; OptiCAM improvements; config docs |
| Multilabel CLI | Parity for bnnr train / analyze workflows |
| Colab | Analyze-only notebook linked from README |
| Batch reports | Run analyze across many checkpoints/configs (foundation for compare) |
Q4 2026 — depth + scaling
| Planned | Description |
|---|---|
| Compare runs | Side-by-side analyze for two checkpoints (compliance / audit use cases) |
| v1.0 criteria | Public stability criteria; reduce “beta” friction on PyPI when met |
| GPU speedups | Faster analyze runtime; caching and better defaults |
| More report templates | Stakeholder-ready summary pages and comparisons |
| HF Hub analyze | Load pretrained weights via Hub URI where feasible |
2027 H1 (gated)
| Planned | Description |
|---|---|
| Detection analyze | Heatmaps + failure buckets for detection — only after classification analyze adoption and clear community demand |
| CI integration | Optional GitHub Actions step producing analyze HTML as artifact |
See detection.md for current train-only detection scope.