How pengyizhang/slimyolov3 works
Project Type and Purpose
• Type: Machine-learning training & inference pipeline, combining Darknet (C) and PyTorch (Python) components.
• Purpose: Learn a sparse (“slim”) YOLOv3 object detector via channel pruning, deployable on resource-constrained UAVs.
Main Architectural Components
A. Data Management
– VisDrone2019 Dataset (images, labels, train/val lists) in VisDrone2019/
– Data‐loader configurations in VisDrone2019/drone.data and cfg/*.cfg
B. Configuration Layer
– Unpruned & pruned network definitions in cfg/00-unpruned, cfg/1iter-pruned, …
– YOLOv3 model configs (yolov3.cfg, yolov3-spp3.cfg, yolov3-tiny.cfg)
C. Training Engines
1. Darknet C Binary
– Normal training: ./darknet detector train ...
– Fine‐tuning pruned networks: ./darknet detector train cfg/prune_*.cfg ...
2. PyTorch Scripts (ultralytics/yolov3)
– Sparsity training: yolov3/train_drone.py with L1‐regularization
– Pruning: yolov3/prune.py and root prune.py (collect channel‐scaling factors, prune)
D. Pruning Workflow
– Input: sparsity-trained weights + config
– Process: channel selection & config rewriting → pruned CFGs + pruned weight files
E. Evaluation & Results
– Darknet inference: darknet detector valid …
– PyTorch inference: yolov3/test_drone.py, test_single_image.py
– Outputs: accuracy metrics, demo video (results/demo.mp4), images directory
Relationships & Data Flow
• Data feeds into both Darknet and PyTorch training engines.
• Each engine reads .cfg and .data files, emits weight files.
• Sparsity training outputs “sparsity.weights” → fed into pruning scripts.
• Pruning scripts read granular cfgs, produce pruned cfgs in cfg/** and pruned‐weights.
• Fine-tuning uses pruned cfgs + pruned weights → final pruned weights in backup directories.
• Evaluation modules consume final weights + configs → produce metrics, demo.mp4.
Architectural Patterns & Design Principles
• Pipeline pattern: a linear sequence (Train → Sparsity → Prune → Fine-tune → Evaluate)
• Config-driven design: behavior controlled by .cfg files, parameters for sparsity, prune ratios
• Modular separation: C‐based Darknet vs Python/PyTorch modules, clear interfacing via configs/weights
• Extensibility: new pruning ratios via directory cfg/***, new datasets via VisDrone2019/*
Technologies & Frameworks
• Darknet (C, GPU accelerated) for core YOLO training/inference
• PyTorch (Python ≥1.0), ultralytics/yolov3 repo for sparsity training & Python-based pruning
• Python scripts: prune.py, sparsity.py (root and yolov3/)
• Data/config files: .data, .cfg, label lists