Skip to content

README.md

Latest commit

 

History

History
191 lines (142 loc) · 4.44 KB

README.md

File metadata and controls

191 lines (142 loc) · 4.44 KB

CropEdgeAI 🌱

Python License PyTorch YOLO

CropEdgeAI is a Computer Vision pipeline for crop and weed detection, with tools for dataset processing, data augmentation, hyperparameter optimization, and NCNN export for edge deployment.

Attribution

This project uses data from:

  • Weed-crop dataset in precision agriculture by Upadhyay et al., North Dakota State University
  • Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0)
  • DOI: 10.17632/mthv4ppwyw.2 Please check NOTICE for further details.

Table of Contents

Features

  • Data loading, validation, splitting, augmentation
  • EDA: dataset statistics and visualizations
  • HPO with Optuna
  • YOLO v8/v9/v10/v11/v12 training and benchmarking
  • NCNN export for edge devices
  • Batch inference with stats export
  • Google Drive backup

Supported Crops

Black bean, Canola, Corn, Field pea, Flax, Lentil, Soybean, Sugar beet

Installation

  • Python 3.11 or 3.12
  • CUDA GPU (optional)
  • 40GB+ disk space
git clone https://github.com/manueljesus/cropedgeai.git
cd cropedgeai
pip install uv
uv sync

For development: uv sync --group dev

Quick Start

1. Data Preparation

from cropedgeai.dataset import WeedCropDatasetLoader

loader = WeedCropDatasetLoader("path/to/weed-crop-dataset")
dataset = loader()

print(f"Loaded {len(dataset)} annotations")
print(f"Classes: {dataset['class_name'].unique()}")

2. Exploratory Data Analysis

from cropedgeai.eda import WeedCropEDA, WeedCropVisualizer

eda = WeedCropEDA(dataset, dataset_name="My Dataset")
stats = eda.dataset_distribution()

visualizer = WeedCropVisualizer(eda)
visualizer.plot_class_balance()
visualizer.plot_spatial_distribution()

3. Dataset Splitting & Augmentation

from cropedgeai.dataset import DatasetSplitter, DatasetAugmenter

splitter = DatasetSplitter(dataset, validation_size=0.15, test_size=0.15, output_dir="./processed", organize_files=True)
train, val, test = splitter()

augmenter = DatasetAugmenter(train, "./augmented", "./processed")
augmented_dataset = augmenter(
    run_overlay=True,
    run_closeup=True,
    overlay_images=1000,
    weeds_per_image=5
)

4. Model Training & Benchmarking

from cropedgeai.experiments import YOLOExperiment

experiment = YOLOExperiment("config/baseline_experiment.yaml")
results = experiment()

5. Hyperparameter Optimization

from cropedgeai.experiments import YOLOHyperparameterOptimizer

optimizer = YOLOHyperparameterOptimizer("config/hpo_yolo11n.yaml")
hpo_results = optimizer()

6. Edge Deployment Optimization

from cropedgeai.experiments import YOLONCNNHyperparameterOptimizer

ncnn_optimizer = YOLONCNNHyperparameterOptimizer("config/ncnn_hpo.yaml")
results = ncnn_optimizer()

ncnn_path = ncnn_optimizer.export_ncnn(best_size) # Check results to confirm `best_size`

7. Batch Inference

cropedgeai model.pt input_folder/ output_folder/ \
  --conf-threshold 0.35 \
  --img-size 640 \
  --stats-csv detections.csv
from cropedgeai.inference import YOLOBatchProcessor

processor = YOLOBatchProcessor(
    model_path="best.pt",
    input_folder="test_images/",
    output_folder="results/",
    stats_csv_file="detection_stats.csv",
    conf_threshold=0.35
)
processor()

Configuration

Experiment

name: "My_Experiment"
dataset_yaml: "data/dataset.yaml"
models_to_test:
  yolo11n: "yolo11n.pt"
  yolo11s: "yolo11s.pt"
training_config:
  epochs: 100
  batch: 32
  imgsz: 640
  patience: 20

HPO

name: "crop_hpo"
dataset_yaml: "data/dataset.yaml"
base_model_path: "yolo11n.pt"
hyperparameter_search_space:
  lr0:
    type: "float"
    min: 0.0001
    max: 0.01
    log: true
  mosaic:
    type: "float"
    min: 0.0
    max: 1.0

CLI

cropedgeai model.pt input/ output/
cropedgeai model.ncnn input/ output/ --conf-threshold 0.4 --img-size 416
cropedgeai --help