Food-11 Image Classification

A food image classification project using transfer learning with ResNet18 and semi-supervised learning, classifying images into 11 food categories.

Dataset

• Source: Food-11 (Kaggle)
• Size: ~16,000 images
• Classes: Bread, Dairy, Dessert, Egg, Fried Food, Meat, Noodles/Pasta, Rice, Seafood, Soup, Vegetable/Fruit
• Splits: Labeled training / Unlabeled training / Validation / Test

⚠️ Download the dataset from Kaggle and extract it into the food-11/ directory at the project root. See DATA.md for detailed setup instructions and class definitions.

food-11/
├── training/
│   ├── labeled/      # Labeled training data, organized into folders 00-10
│   └── unlabeled/    # Unlabeled data for semi-supervised learning
├── validation/       # Validation set
└── testing/          # Test set

Project Structure

food_classification/
├── model_utils/              # Model-related code
│   ├── data.py               # Dataset classes, data loading & augmentation
│   ├── model.py              # Model definitions & pretrained model loading
│   └── train.py              # Training/validation loop with semi-supervised support
├── notebooks/                # Jupyter Notebooks
│   ├── 01_food_classification.ipynb
│   └── acc.png               # Training accuracy curve
├── figures/                  # Generated visualizations
│   ├── acc.png               # Training accuracy curve
│   └── prediction_samples.png
├── model_save/               # Saved model weights (not version-controlled)
├── food-11/                  # Full dataset (download separately)
├── food-11_sample/           # Small sample dataset (for testing)
├── main.py                   # Main training script
├── requirements.txt
├── DATA.md                   # Dataset documentation
├── LICENSE
└── README.md

Getting Started

1. Install Dependencies

pip install -r requirements.txt

2. Prepare Data

Download the Food-11 dataset from Kaggle and extract it to the project root:

# Expected structure: food-11/training/labeled/00/, 01/, ..., 10/

3. Train the Model

Option 1: Command line

python main.py

Option 2: Jupyter Notebook (recommended)

jupyter notebook notebooks/01_food_classification.ipynb

The notebook covers the full pipeline: data loading → model initialization → training → prediction visualization → accuracy curves.

Model & Methods

Multi-Architecture Support

Switch between pretrained models with a single line:

model, size = initialize_model("resnet18", num_classes=11, use_pretrained=True)
# model, size = initialize_model("vgg", num_classes=11, use_pretrained=True)
# model, size = initialize_model("densenet", num_classes=11, use_pretrained=True)
# model, size = initialize_model("MyModel", num_classes=11)  # Custom CNN

Supported architectures: ResNet18/50, VGG11, DenseNet121, GoogLeNet, AlexNet, SqueezeNet, Inception v3, and a custom CNN.

Training Configuration

Config	Value
Transfer Strategy	Fine-tuning (all parameters)
Data Augmentation	RandomResizedCrop + RandomHorizontalFlip + AutoAugment
Optimizer	AdamW (lr=5e-4, weight_decay=1e-4)
Loss Function	CrossEntropyLoss
Batch Size	32
Epochs	50

Semi-Supervised Learning

Once validation accuracy exceeds 70%, semi-supervised learning is automatically enabled: the current model predicts labels for the unlabeled data, and samples with confidence > 99% are added to the training set as pseudo-labeled data.

Results

Metric	Value
Best Validation Accuracy	~85%

The best-performing model on the validation set is automatically saved to model_save/model.pth during training.

Code Overview

• model_utils/data.py — Custom foodDataset class for loading labeled/unlabeled data with augmentation; noLabDataset class for generating pseudo-labeled datasets.
• model_utils/model.py — Wraps multiple pretrained models with classifier head replacement; supports both linear probing and fine-tuning.
• model_utils/train.py — Training loop with integrated semi-supervised learning and automatic loss/accuracy curve plotting.
• main.py — Configures hyperparameters and launches training.

Notes

• If GPU memory is insufficient, reduce batchSize (e.g., to 16) or use the smaller food-11_sample/ dataset for testing
• Semi-supervised learning thresholds can be adjusted in main.py (acc_thres and conf_thres)

License

MIT