Project Scripts & Usage Guide

Scripts Overview

• data/data_prep.py: Handles data splitting (train/val/test) and computes class weights to handle class imbalance. It generates indices in data/splits/.
• data/extract_features.py: Runs a pre-trained ResNet18 (without the final classification layer) on the dataset to extract 512-dimensional feature vectors for every image. Saves these features to .npz files in features/.
• CNN/train.py: A sample script that trains a simple Convolutional Neural Network (CNN) from scratch using the raw image data and the generated splits.
• Random_Forest/train.py: Trains a scikit-learn RandomForest on the precomputed feature .npz files and saves metrics/model artifacts.

Environment Setup

1. Ensure you have Conda installed.
2. Create the environment using the provided environment.yml:

   conda env create -f environment.yml

3. Activate the environment:

   conda activate plant_env

Alternative Setup (pip / Google Colab)

If you do not have Conda (e.g., on Google Colab), you can install the required packages using pip:

pip install torch torchvision numpy scikit-learn pandas matplotlib pillow tqdm

Data Preparation

1. Extract Raw Data

If the data is not yet extracted, use the following command to extract the tarball:

tar -xvf data/house_plant_species_raw/house_plant_species.tar -C data/

2. Splits & Feature Extraction

Note: These steps are likely already completed. Check data/splits and features/.

If you need to regenerate splits or features:

# Generate splits
python data/data_prep.py --data_root data/house_plant_species --splits_dir data/splits

# Extract features (requires splits to exist)
python data/extract_features.py --data_root data/house_plant_species --splits_dir data/splits --out_dir features/

# Train Random Forest on features
python Random_Forest/train.py --features_dir features --out_dir Random_Forest/output

Since feature extraction is already complete, you do not need to load raw images or perform computationally expensive preprocessing. You can load the pre-computed feature vectors directly.

Step-by-Step Guide

1. Create a new directory for your model (e.g., SVM/, RandomForest/).
2. Create a training script (e.g., SVM/train_svm.py).
3. Load the data: Use numpy.load to read the .npz files from the features/ directory.
4. Train & Evaluate: Use Scikit-Learn or other libraries to train on the loaded arrays.

Example Code Snippet

Here is a template for a new training script (e.g., SVM/train.py):

import sys
import numpy as np
from pathlib import Path
from sklearn.svm import SVC
from sklearn.metrics import accuracy_score

# 1. Setup Paths
PROJECT_ROOT = Path(__file__).resolve().parents[1]
FEATURES_DIR = PROJECT_ROOT / "features"

def load_features(split_name):
    """Load X and y from .npz file."""
    file_path = FEATURES_DIR / f"{split_name}_features.npz"
    data = np.load(file_path)
    return data['X'], data['y']

def main():
    print("Loading features...")
    X_train, y_train = load_features("train")
    X_val, y_val = load_features("val")
    X_test, y_test = load_features("test")

    print(f"Train shape: {X_train.shape}")
    
    # 2. Initialize Model
    # Example: Support Vector Machine
    clf = SVC(kernel='linear', C=1.0, random_state=42)
    
    # 3. Train
    print("Training SVM...")
    clf.fit(X_train, y_train)
    
    # 4. Evaluate
    print("Evaluating...")
    val_preds = clf.predict(X_val)
    val_acc = accuracy_score(y_val, val_preds)
    print(f"Validation Accuracy: {val_acc:.4f}")

    test_preds = clf.predict(X_test)
    test_acc = accuracy_score(y_test, test_preds)
    print(f"Test Accuracy: {test_acc:.4f}")

if __name__ == "__main__":
    main()