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)

AreaWhat exists today
Analyzebnnr analyze + portable HTML/JSON for classification and multilabel; sample report
TrainingCLI train / demo / quickstart; presets; live dashboard
XAI augICD, AICD, branch search; plugin_icd.md for custom loops
DetectionTrain API + bbox augs; Ultralytics adapter; detection.md — no analyze for detection yet
BenchmarksCIFAR-10 demo CNN (3 seeds): no aug vs RandAugment vs BNNR — benchmarks.md
Integrationspytorch-grad-cam loop, Ultralytics quickstart — integrations.md
Torchvision analyzePython example: torchvision_analyze_cifar10.py

Q2 2026 (June–August) — credibility + analyze adoption

PlannedDescription
Analyze-first docsREADME and getting_started.md: try analyze before train; githack sample links
Augmentation benchmarkResNet18 / Imagewoof (fine-grained, low-data, from-scratch): 5 seeds; baselines RandAugment + TrivialAugment; reproducible script benchmarks/reproduce_imagewoof.sh (free-T4 friendly)
Torchvision → analyzeProminent golden path: pretrained classifier → HTML report; runnable example in README
Repo discoverabilityGitHub description, website link, topics (maintainer ops)
Contributor templatesIssue forms: benchmark proposal, “Who uses BNNR” showcase
Upstream docsOfficial link in Ultralytics docs (after PR merge); grad-cam ecosystem follow-up
HuggingFace SpaceMinimal Gradio demo (sample report or analyze workflow)

Q3 2026 (September–November) — analysis depth + ecosystem

PlannedDescription
HuggingFace modelsFirst 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 methodsBetter stability + more methods; OptiCAM improvements; config docs
Multilabel CLIParity for bnnr train / analyze workflows
ColabAnalyze-only notebook linked from README
Batch reportsRun analyze across many checkpoints/configs (foundation for compare)

Q4 2026 — depth + scaling

PlannedDescription
Compare runsSide-by-side analyze for two checkpoints (compliance / audit use cases)
v1.0 criteriaPublic stability criteria; reduce “beta” friction on PyPI when met
GPU speedupsFaster analyze runtime; caching and better defaults
More report templatesStakeholder-ready summary pages and comparisons
HF Hub analyzeLoad pretrained weights via Hub URI where feasible

2027 H1 (gated)

PlannedDescription
Detection analyzeHeatmaps + failure buckets for detection — only after classification analyze adoption and clear community demand
CI integrationOptional GitHub Actions step producing analyze HTML as artifact

See detection.md for current train-only detection scope.