Plant Disease Classification 🌿

This project implements deep learning models for plant disease classification using leaf images. We experiment with several model architectures:

• Multi-Layer Perceptron (MLP)
• Convolutional Neural Networks (CNN)
• Vision Transformers (ViT)

---

Environment Setup

1. Clone the repository:

git clone https://github.com/lolyhop/plant-disease-classification.git
cd plant-disease-classification

2. Install dependencies:

pip install -r requirements.txt

3. Download the dataset:

bash scripts/download_raw_data.sh

---

Repository Structure

.
├── configs/                  # Model configurations
├── data/                     # Dataset and logs
├── docs/                     # Images and documentation
├── front/                    # Streamlit frontend
├── notebooks/                # Jupyter notebooks for exploration
├── scripts/                  # Helper scripts
├── src/                      # Code for models, training, and utilities
├── train.py                  # Training script
├── inference.py              # Inference script
├── requirements.txt
└── README.md

---

Run Experiments

You can run training and inference for any model using a config file.

Training:

python train.py --config configs/<config_name>.yaml

Examples:

python train.py --config configs/resnet.yaml
python train.py --config configs/efficientnet.yaml
python train.py --config configs/vit_2021.yaml

Inference:

python inference.py --config configs/<config_name>.yaml

Examples:

python inference.py --config configs/mlp.yaml
python inference.py --config configs/densenet.yaml
python inference.py --config configs/t2t_vit.yaml

This approach works for all models and keeps commands consistent.

---

Model API Deployment

1. Build the Docker image

cd plant-disease-classification
docker build -f deploy/Dockerfile -t plant-classifier .

2. Run the API container

Mount your local data directory and provide model config and weights:

cd plant-disease-classification
docker run -p 8000:8000 \
  -v $(pwd)/data:/app/data \
  -e MODEL_CONFIG_PATH=/app/configs/<your_model>.yaml \
  -e MODEL_WEIGHTS_PATH=/app/data/logs/<your_model>/best_model.pt \
  plant-classifier

• -p 8000:8000 exposes the API on port 8000.
• -v $(pwd)/data:/app/data allows the container to access model weights.
• MODEL_CONFIG_PATH and MODEL_WEIGHTS_PATH must point to the YAML config and .pt file inside the container.

3. Test the API

OpenAPI JSON: http://localhost:8000/openapi.json

Send a POST request to /predict with a leaf image:

curl -X POST "http://localhost:8000/predict" \
  -F "file=@path_to_leaf_image.jpg"

The API returns a JSON dictionary with class probabilities:

{
  "Tomato___Early_blight": 0.92,
  "Tomato___Leaf_Mold": 0.05,
  "Tomato___healthy": 0.02,
  ...
}

---

Run Frontend

Navigate to the frontend directory and set environment variables:

cd front/app

export MODEL_CONFIG_PATH=../../configs/<config_name>.yaml
export MODEL_WEIGHTS_PATH="weights/best_model.pt"

python main.py

Available configs: deep_mlp.yaml, densenet.yaml, efficientnet.yaml, mlp.yaml, resnet.yaml, t2t_vit.yaml, vit_2021.yaml

---

Model Performance

Model	Accuracy	Precision	Recall	F1-score
DenseNet-121	0.9961	0.9962	0.9960	0.9960
ResNet-18	0.9959	0.9958	0.9959	0.9958
EfficientNet-B0	0.9950	0.9952	0.9947	0.9948
Tokens-to-Token ViT (T2T-ViT)	0.9929	0.9930	0.9929	0.9929
Vision Transformer (ViT)	0.9909	0.9911	0.9908	0.9909
Deep MLP	0.7620	0.7755	0.7618	0.7601
MLP	0.0552	0.0430	0.0522	0.0263

Top performing models are CNNs and Transformer-based models. Simple MLPs perform poorly, highlighting the importance of spatial feature extraction.

---

Tech Stack

• Python 3.10+
• PyTorch
• Albumentations
• Torchvision