Math156-Rice-Classification

A deep learning project classifying rice grain images into five varieties using a custom Convolutional Neural Network (CNN) with hyperparameter optimization.

Dataset

Rice Image Dataset: 75,000 images across 5 rice varieties (Arborio, Basmati, Ipsala, Jasmine, Karacadag)
Download from Kaggle

• Training: 60%
• Validation: 20%
• Testing: 20%
• Original resolution: 250×250 pixels

Note: Update dataset path in the code:

ROOT = '/path/to/your/Rice_Image_Dataset'

Installation

Clone the repository:

git clone https://github.com/FarrelMax/Math156-Rice-Classification.git
cd Math156-Rice-Classification

Install dependencies:

pip install torch torchvision numpy matplotlib scikit-learn pandas seaborn tqdm optuna torchsummary

For Google Colab users, mount your Drive where the dataset is stored.

Model Architecture

Custom CNN (~54K trainable parameters):

1. Conv2D → ReLU → MaxPool
2. Conv2D → ReLU → MaxPool
3. AdaptiveAvgPool → Flatten
4. Fully connected layer → 5-class output

• Supports tunable hidden size and learning rate.

Key Features

• Data augmentation: random flips, rotation, color jitter
• Dataset-specific normalization
• Hyperparameter tuning with Optuna
• Evaluation: confusion matrix, classification report, PCA, and K-Means clustering

Inference

To use a trained model for inference:

import torch
from rice_classification import CNN  # or the appropriate module

# Set device
device = torch.device(
    "mps" if torch.backends.mps.is_available()
    else "cuda" if torch.cuda.is_available()
    else "cpu"
)

# Load trained model
model = CNN(num_classes=5, hidden_size=32)
model.load_state_dict(torch.load('rice_cnn_best_model.pth', map_location=device))
model.to(device)
model.eval()

## Results

| Metric                   | Value             |
| ------------------------ | ----------------- |
| Test Accuracy            | 98.8%             |
| Best Validation Accuracy | 99.10% (epoch 14) |

Seed = 42 for reproducibility.

## Key Hyperparameters

```python
BATCH_SIZE = 64
NUM_EPOCHS = 20
LEARNING_RATE = 0.001
HIDDEN_SIZE = 32
NUM_TRIALS = 20  # Optuna

Tuned parameters:

• Learning rate: 1e-4 → 1e-2 (log scale)
• Hidden layer size: 16, 32, 64
• Weight decay: 1e-6 → 1e-3 (log scale)

Citations

Dataset

Koklu, M. (2021). Rice Image Dataset. Kaggle.

Core Methods & Theory

LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324.

Libraries & Frameworks

• PyTorch Documentation — Data utilities: torch.utils.data
• GeeksforGeeks tutorial: "Building a Convolutional Neural Network Using PyTorch"
• GeeksforGeeks tutorial: "Hyperparameter Tuning with Optuna in PyTorch"

Implementation Resources

• GeeksforGeeks: "How Do You Use PyTorch's Dataset and DataLoader Classes for Custom Data?"
• Hosseini, G. (2023). "Beginner Tutorial: Image Classification Using PyTorch." Medium.
• Rath, S. R. (2021). "PyTorch ImageFolder for Training CNN Models." DebuggerCafe.

---

Authors

Sophia Palomares
Farrel Gomargana
Diana Tolu
MATH 156 – UCLA
November 26, 2025

---

For complete methodology and analysis, refer to the project report.