README.md

metadata

title: Food-101 Classifier
emoji: 🍔
colorFrom: yellow
colorTo: red
sdk: gradio
app_file: scripts/app.py

🍽️ Food-101 Image Classification with EfficientNetV2-S and PyTorch Lightning

This repository contains the code for an end-to-end deep learning project to classify 101 food categories from the challenging Food-101 dataset. The project demonstrates a systematic approach to model selection, fine-tuning, and hyperparameter optimization, achieving a final validation accuracy of 85.4% on the full dataset.

The entire training and evaluation pipeline is built using modern, reproducible practices with PyTorch Lightning.

---

📑 Table of Contents

• 🍽️ Food-101 Image Classification with EfficientNetV2-S and PyTorch Lightning
  • 📑 Table of Contents
  • 🎯 Project Highlights
  • 💡 Real-World Applications
  • 🧫 Experimental Results
  • 📊 Final Results
  • 🔬 Performance Analysis and Error Diagnosis
    • 🍤 Lowest-Performing Classes
    • Root Cause Analysis of Misclassifications
    • Future Work
  • 🧪 Methodology and Experimental Process
  • 📁 Repository Structure
  • 🚀 Getting Started
    • Prerequisites
    • Installation
    • Usage
    • 💻 Technologies Used

---

🎯 Project Highlights

• High-Performance Model ⚡: Utilizes a pre-trained EfficientNetV2-S, selected for its excellent balance of accuracy and computational efficiency suitable for potential edge deployment.
• Reproducible Pipeline 🔄: Encapsulates the entire workflow—from data loading to training and evaluation—in a clean and organized LightningModule and DataModule.
• Efficient Experimentation ⏱️: Overcame hardware limitations by implementing dataset subsetting for rapid prototyping.
• Advanced Fine-Tuning 🛠️: Implemented a robust fine-tuning strategy, unfreezing the final three blocks of the feature extractor and using the Adam optimizer with a CosineAnnealingLR scheduler for stable convergence.
• In-Depth Analysis 🔎: Went beyond simple accuracy by calculating and logging per-class F1-scores and accuracies, enabling a deep dive into the model's strengths and weaknesses.
• Live Deployment 📺: The final model is deployed and accessible as an interactive Gradio web application on Hugging Face Spaces.

---

💡 Real-World Applications

Beyond being a technical challenge, this food classification model serves as a foundation for numerous real-world applications in health, hospitality, and smart home technology.

• Health and Nutrition Tracking

  • Automated Calorie Counting: Users can snap a photo of their meal, and an app can automatically identify each food item to provide an instant estimate of calories, macros, and other nutritional information.
  • Dietary Management: Assists individuals with allergies or specific dietary needs (e.g., diabetes, gluten-free) by helping them identify and log their food intake accurately.

• Restaurant and Hospitality Tech

  • Self-Checkout Systems: In cafeterias or quick-service restaurants, a camera-based system could identify all items on a tray to automate the billing process, reducing queues and improving efficiency.
    • Interactive Menus: Allow diners to point their phone at a dish to get more information, such as ingredients, allergen warnings, or customer reviews.

• Smart Home and Appliances

  • Smart Refrigerators: A fridge equipped with a camera could identify leftover dishes, suggest recipes based on available food, and help track food spoilage to reduce waste.

---

🧫 Experimental Results

This project followed an iterative approach. The table below summarizes the key experiments and their outcomes, showing the progression from the initial baseline to the final model.

Model	Training Strategy	Data %	Key Hyperparameters	Final Val Accuracy
EfficientNet-B2	Simple fine-tune (last block)	50%	lr=1e-4	~64%
EfficientNet-B2	Unfreeze last 3 blocks	50%	lr=1e-3	82.0%
EfficientNet-B2	Two-Stage Fine-Tuning	50%	lr1=1e-3, lr2=1e-5	Performance Degraded
EfficientNetV2-S	Unfreeze last 3 blocks	50%	lr=1e-4 (Tuned)	82.4%
EfficientNetV2-S	Unfreeze last 3 blocks and more advanced transforms	50%	lr=1e-4 (Tuned)	~82.4% Pretty much the same Performance
EfficientNetV2-S	Unfreeze last 3 blocks	100%	lr=1e-4 (Tuned)	85.4%

---

📊 Final Results

After systematically iterating on model architecture and hyperparameters, the final model achieved the following performance on the full Food-101 validation set:

Metric	Score
Validation Accuracy	85.4%

Here is the Confusion Matrix on the Test set. (you can find this plot in the assets section)

This model is deployed and accessible as an interactive Gradio web application on Hugging Face Spaces.

Check out my Food101 Gradio Demo.

---

🔬 Performance Analysis and Error Diagnosis

Beyond the aggregate accuracy, a per-class analysis was conducted to identify the model's specific limitations and diagnose the root causes of misclassifications.

The model performed exceptionally well on many classes but struggled with a distinct set of categories, primarily due to visual ambiguity and high variability in appearance.

🍤 Lowest-Performing Classes

The following five classes had the lowest validation accuracy:

Class Name	Index	Validation Accuracy
shrimp_and_grits	93	44.0%
ravioli	77	59.2%
apple_pie	0	61.6%
huevos_rancheros	56	63.2%
falafel	36	63.6%

Root Cause Analysis of Misclassifications

• High Intra-Class Variation: The model struggled with dishes that have no single, consistent appearance.
• Fine-Grained Confusion: Errors occurred between visually similar classes like ravioli vs. dumplings.
• Ambiguous Features: Foods like falafel resemble many small fried dishes, making classification tricky.

Future Work

Improvements could include:

• Detailed confusion matrix analysis 🔍
• More aggressive data augmentation 📈
• Larger architectures for fine-grained recognition 🏋️
• Training for longer 🏋️

---

🧪 Methodology and Experimental Process

Steps taken in the project:

1. Baseline Establishment 🏁 – EfficientNet-B2 achieved ~64%.
2. Architecture Selection 🏗️ – EfficientNetV2-S chosen for balance of accuracy and size.
3. Transforms Selection 🎨 – TrivialAugmentWide + RandomResizedCrop, RandAugment, etc.
4. Fine-Tuning Strategy 🔧 – Final 3 blocks unfrozen for training.
5. Final Model Training 🏆 – Full dataset, Adam, CosineAnnealingLR, EarlyStopping → 85.4%.

---

📁 Repository Structure

food-101-classification/
├── data/
├── logs/
├── scripts/
│   ├── main.py
│   ├── models.py
│   ├── class_names.py
│   ├── app.py
│   └── prepare_data.py
├── .gitignore
├── requirements.txt
└── README.md

---

🚀 Getting Started

Prerequisites

• Python 3.10+ 🐍
• PyTorch 🔥
• CUDA-enabled GPU (recommended) 🎮

Installation

1. Clone the repository:

   git clone https://github.com/Deathshot78/Food101-Classification
   cd Food101-Classification
   

2. Install the dependencies:

   pip install -r requirements.txt
   

Usage

Run training with a subset for quick testing:

python main.py

💻 Technologies Used

• Python

• PyTorch

• PyTorch Lightning

• TorchMetrics

• Gradio

• Matplotlib & Seaborn