PlantAi — Plant Disease Detection (Model Training)

PlantAi is a CNN trained on PlantVillage to classify leaf images as healthy or diseased. This repository provides PyTorch training code and export tools (ONNX / TorchScript / TFLite) for deployment.

Overview

• Training-first: Config-driven PyTorch pipeline (dataloader, augmentation, logging, plotting).
• Exports: ONNX / TorchScript / TFLite for downstream deployment.
• Lightweight inference: Designed for CPU-friendly inference after export.

---

Requirements & Installation

# from repo root
pip install -r requirements.txt

Python 3.9+ recommended. Use a virtualenv/conda if preferred.

---

Training (PyTorch)

1. Dataset placement Download PlantVillage and place under model/data/<DATASET_NAME>/ in class‑subfolders (ImageFolder layout).

   • GitHub: https://github.com/spMohanty/PlantVillage-Dataset
   • Kaggle: https://www.kaggle.com/datasets/nishantvalvi1504/plantvillagecomplete

2. Config Edit model/config/plantai_resnet_18.yaml:

   • DATASET.PATH: model/data
   • DATASET.NAME: <DATASET_NAME>
   • OUTPUT_DIR: log/plant_ai_resnet18 (or your choice)

3. Run

cd model
python train.py --config_file config/plantai_resnet_18.yaml

Checkpoints/plots are written under model/log/....

---

Model Export

Ensure MODEL.WEIGHTS in the YAML points to your trained best_model.pth.

cd model

# ONNX
python export/export_to_onnx.py --config_file config/plantai_resnet_18.yaml

# TorchScript (Lite)
python export/export_to_torchscript.py --config_file config/plantai_resnet_18.yaml

# TensorFlow Lite
python export/export_to_tf_light.py --config_file config/plantai_resnet_18.yaml --output plantai_model.tflite

---

Dataset (PlantVillage)

• ~54,305 images of healthy/diseased leaves captured on plain backgrounds.
• 38 classes total (26 diseases + healthy across common crops).
• ~17 crop types (e.g., apple, grape, corn, tomato, potato, pepper, etc.).
• Public, widely used for supervised learning; ideal for compact models and mobile inference.

Source: https://github.com/spMohanty/PlantVillage-Dataset

Sample grid

---

Results (training snapshots)

Evaluation (test set)

• Accuracy: 93.14%
• Precision (weighted): 92.90%
• Recall (weighted): 93.14%
• F1 Score (weighted): 92.50%

Pretrained Weights

ResNet-18 (PlantVillage) — Download weights

---

Acknowledgments

• Dataset: PlantVillage — https://github.com/spMohanty/PlantVillage-Dataset