How meteorsee/mknet works
MKNet is a lightweight YOLO-based object-detection model for resource-constrained and assistive-vision use. Its principal workflow is to train a configured model on image data, then validate or run inference to obtain detections. The repository includes training and validation scripts, model-building blocks, and ablation configurations; the README also documents CLI inference and COCO-style evaluation. Available source excerpts do not establish how the YAML variants assemble the blocks, and the documented inference and COCO-evaluation scripts are absent from the supplied tree, so those links are omitted.
Main entry points: train.py Custom training · val_model.py Validation · ablation Ablation variants · comparison_results.csv Reported metrics
Components 15
Model architecture
- Ablation variants
ablation - MobileNetV3 blocks
models/MobileNetV3.py - SPD-DW downsampling
models/MobileNetV3.py - CSPCAM context
models/MobileNetV3.py - AKMix partial convolution
models/MobileNetV3.py - GSConv fusion blocks
models/GSConv.py
Training and evaluation
- Custom training
scripts/train.py - Validation
scripts/val_model.py - CLI inference
- COCO evaluation
- Reported metrics
results/comparison_results.csv - Ablation metrics
results/ablation_results.csv
Other components
- Engineer
- Detection dataset
- Ultralytics YOLO
Connections 11
- Engineer → Custom training: starts training
- Custom training → Ultralytics YOLO: calls train and val
- Detection dataset → Ultralytics YOLO: supplies samples
- Engineer → Validation: starts validation
- Validation → Ultralytics YOLO: calls validation
- Engineer → CLI inference: requests detections
- CLI inference → Ultralytics YOLO: runs prediction
- Engineer → COCO evaluation: requests evaluation
- COCO evaluation → Ultralytics YOLO: runs validation
- Ultralytics YOLO → Reported metrics: produces metrics
- Ablation variants → Ablation metrics: compared by