Cinnamon Quill Grade Classification (PyTorch)

This project trains deep learning models to classify cinnamon quill images into four quality grades:

• Alba
• C4
• C5
• C5 Special

The workflow includes preprocessing, augmentation, two CNN architectures, transfer learning experiments, TensorBoard monitoring, evaluation, and saving trained models in .pth format.

---

Project Structure

project/
│── datasets/
│ └── cinnamon/
│ ├── Alba/
│ ├── C4/
│ ├── C5/
│ └── C5 Special/
│
│── checkpoints/
│ ├── saved_model_pretrained.pth
│ └── saved_model_scratch.pth
│
│── utils.py
│── train.py
│── test.py
│── predict.py
│── README.md

---

1. Dataset Preparation

Download dataset

Place cinnamon images in labeled class folders.

Preprocessing

All images are resized and normalized:

• Resize to 224×224
• Normalize using ImageNet mean/std
• Convert to tensors

Data Augmentation

Used during training:

• Random rotations
• Horizontal & vertical flips
• Random cropping
• Optional color jitter

These help models generalize better and deal with class imbalance.

---

2. Dataset Splitting

Dataset is split using:

• 70% Training
• 15% Validation
• 15% Test

Using random_split from PyTorch.

---

3. CNN Architectures

Two network architectures were created:

ResNet18

• With transfer learning
• Without transfer learning (trained from scratch)

VGG16

• With transfer learning
• Without transfer learning

Toggle transfer learning:

use_pretrained = True    # Transfer learning
use_pretrained = False   # Train from scratch

Final Classifier

The final classifier layers are adapted for 4 output classes: Alba, C4, C5, and C5 Special.

---

4. Training

Both ResNet18 and VGG16 models were trained using the following configurations:

Optimizers

• SGD
• Adam

Loss Function

• CrossEntropyLoss

Additional Training Details

• WeightedRandomSampler to reduce class imbalance
• Batch size = 8
• Learning rate scheduler
• Dropout added to VGG classifier
• Training and validation logs tracked in TensorBoard:

tensorboard --logdir=runs

5. Evaluation

Models are evaluated on the test set using:

• Test accuracy
• Test loss
• F1-score
• Class-wise accuracy
• Confusion matrix

A final summary table is generated:

Model	Optimizer	Loss	Accuracy	F1	Classwise Accuracy
ResNet18	SGD	...	...	...	...
ResNet18	Adam	...	...	...	...
VGG16	SGD	...	...	...	...
VGG16	Adam	...	...	...	...

---

6. Saving Models

Each experiment saves a .pth model checkpoint:

• saved_model_pretrained.pth
• saved_model_scratch.pth

Saved using:

torch.save({
    "state_dict": model.state_dict(),
    "optimizer": optimizer.state_dict()
}, "checkpoints/model_name.pth")

7. Prediction Script

A separate predict.py script loads a trained .pth model and predicts the class of a single input image.

Run it with:

python predict.py

Example Output

Image: datasets/cinnamon/C5/C5 01.JPG
Predicted Grade: C5

---

Completed Requirements Checklist

• Download and preprocess dataset
• Data augmentation
• Train/Val/Test split
• ResNet18 + VGG16 implemented
• Transfer learning ON/OFF
• Train using SGD & Adam
• TensorBoard logging
• Test accuracy and metrics
• Save model in .pth format
• Prediction script

---

Notes

• Imbalanced data may require WeightedRandomSampler.
• Transfer learning typically increases performance.
• Include TensorBoard accuracy / loss plots in your report.

---

Final Goal

Build a reliable cinnamon grading model using deep learning and analyze how:

• architecture choice (ResNet vs. VGG), and
• training strategy (transfer learning vs. training from scratch)

affect final performance.