Skip to content

Releases: KeenForgeAI/KeenForge

Release list

v2.0.1

v2.0.1 Latest
Latest

Choose a tag to compare

View all tags
@lugan113 lugan113 released this 01 Oct 12:28
v2.0.1
1e76932

KeenForge 2.0.0

Released: 2026-10-01 · Branch: main · Tag: v2.0.0

KeenForge 2.0 turns the "label everything, then train" workflow into a closed
cluster-guided active-learning loop. It is still a fully local, offline, no-code
desktop app — your images never leave your machine.

Highlights

  • 🧹 3-layer dataset deduplication — perceptual hash → pixel-level MAE → dihedral
    transforms + ±2px translation alignment. Correctly flags near-duplicates that
    differ only by a 2-pixel shift (MAE 10.14 → 2.57 on a real NEU-DET pair).
  • 🔁 Cluster-guided active learning — cosine clustering picks the most informative
    representatives (typically 15–50 per round). Label those, the model trains in the
    background, then you compare and correct — and the cycle repeats at finer granularity.
  • 🔍 Model Compare Mode — run your trained model on already-labeled images.
    Predictions are cyan dashed boxes with confidence; human labels stay green.
    Edit to accept — un-reviewed predictions can never overwrite ground truth.
  • ⚖️ Bring-your-own weights — resume from any .pt / .pth; transfer / recursive /
    from-scratch strategies.
  • ⚡ Non-blocking training — training runs in a separate process; keep labeling while it runs.
  • 📊 Training summary — per-class bounding-box statistics, per-round mAP, model files.
  • 🌏 Full zh / en UI — switch language at runtime.
  • 🛡️ Robust image I/O — decodes by content, so mislabeled extensions (.png files
    that are really JPEG) just work.
  • 🧠 Model zoo — YOLOv8 n/s/m/l/x, YOLO11 n/s/m/l/x, RT-DETR-l/x, YOLO-World v2,
    Faster-RCNN, RetinaNet, SSD300.

Import / Export

  • Import: COCO JSON, Pascal VOC XML
  • Export: COCO JSON, Pascal VOC XML, YOLO TXT
  • Training output: standard Ultralytics .pt

Install / Run

Run from source (recommended for now)

git clone https://github.com/KeenForgeAI/KeenForge.git
cd KeenForge
pip install -r requirements.txt
python src/main.py

Windows users who build an executable can find the PyInstaller entry point in
entry.py (the spec is kept local).

Install / Run

One-click launcher (recommended, ~5 MB download)

Download KeenForge-2.0.0-source.zip from Releases, unpack it, then:

  • Windows: double-click run_windows.bat
  • Linux / macOS: ./run_unix.sh

The launcher creates a local .venv, installs dependencies automatically on first run
(a few minutes), and starts KeenForge. The heavy PyTorch/Ultralytics stack is fetched on
your machine instead of being bundled, so the download stays small.

Manual (from source)

git clone https://github.com/KeenForgeAI/KeenForge.git
cd KeenForge
pip install -r requirements.txt
python src/main.py

A self-contained Windows executable is possible but large (CUDA-enabled PyTorch pushes
it past 4 GB unpacked); the launcher route is preferred for distribution.

Datasets

Corrected, ready-to-use industrial datasets (deduped, re-split, format-fixed):

GC10-DET · NEU-DET · DeepPCB · PKU-Market-PCB · TXL-PBC · Raccoon
→ https://huggingface.co/KeenForgeAI

Known issues

  • pip install torch on Windows installs the CPU build by default. For NVIDIA GPU
    training/inference, install the CUDA build:
    pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121.
  • Training mAP requires torchmetrics; it is included in requirements.txt.

Roadmap 2.x

  • SAM-powered segmentation (polygon masks)
  • Adaptive fine-tuning tuned for YOLO
  • One-click API deployment
  • pip install support

Citation

@misc{keenforge2026,
  title={KeenForge: No-Code Desktop Tool for Expert-Led Object Detection Model Training},
  author={Lu Gan and Xi Li},
  year={2026},
  howpublished={\url{https://github.com/KeenForgeAI/KeenForge}},
  note={CVPR 2026 Demo Track}
}
Assets 3
Loading

KeenForge v1.0.0

KeenForge v1.0.0

Choose a tag to compare

View all tags
@lugan113 lugan113 released this 04 Aug 06:50
v1.0.0
3cf2a12

🚀 KeenForge v1.0.0 — Forge your own AI vision model, no code needed.

What's included

  • Zero-shot annotation via YOLO-World (type "welding defect" → it finds them)
  • Self-looping active learning (label 15 images → auto-train → model improves)
  • 100% local & offline — your data never leaves your computer
  • COCO / VOC / YOLO format import/export
  • Multiple model support (YOLOv8/11, RT-DETR, Faster-RCNN, SSD)

Downloads

  • KeenForge.exe — Windows 64-bit, ~335 MB

System requirements

  • Windows 10/11
  • 4GB+ RAM (8GB recommended)
  • CPU-only works; GPU (CUDA) recommended for faster training

First launch

  • May take 10-30s to initialize (model loading)
  • Load an image folder, type what to detect, start labeling
Loading