🍛 Nepali Food Classifier

An end-to-end computer vision project that classifies traditional Nepali dishes using transfer learning (ResNet18), compares augmentation strategies, and deploys predictions through Streamlit and FastAPI.

---

Challenges Faced

• Small dataset size
• Class imbalance
• Overfitting in baseline model
• Generalization issues

Solutions:

• Strong augmentation
• Transfer learning
• Fine tuning

---

Streamlit App

Run the web app: View Project

---

FastAPI

POST /predict

---

Project Overview

This project trains two separate models using the same architecture (ResNet18) but different data preprocessing strategies:

Model	Description
Model 0	(Basic Augmentation) Minimal augmentation for baseline performance
Model 1	(Strong Augmentation) Advanced augmentation for better generalization

The goal is to analyze how augmentation impacts:

• Accuracy
• Overfitting
• Generalization
• Robustness

---

Project Structure

project/
│
├── dataset/
│   ├── train/
│   └── test/
│
├── src/
│   ├── dataset.py
│   ├── model.py
│   ├── engine.py
│   ├── predict.py
|   ├── train.py
│   └── streamlit_app
│    
├── models/
└── README.md

---

Dataset

Source: Kaggle Food Dataset (custom Nepali food subset) Data Source

Classes used:

• momo
• dal_bhat
• sel_roti
• chowmein
• thukpa
• chatamari
• yomari

mage format: RGB
Input size: 224 x 224

---

Model

Base architecture:

ResNet18

Transfer learning used with frozen feature extractor and custom classifier head.

---

Data Augmentation Strategies

Model 0 — Basic Augmentation

Used for baseline training:

transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
])

Characteristics:

• No heavy augmentation

• Preserves original data distribution

• Faster convergence

• Higher risk of overfitting

  Model B — Strong Augmentation

Used for better generalization:

transforms.Compose([
    transforms.RandomResizedCrop(224),
    transforms.RandomHorizontalFlip(),
    transforms.RandomRotation(20),
    transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
    transforms.ToTensor(),
])

Characteristics: *More robust to real-world variation *Reduces overfitting *Slightly slower convergence *Better generalization

---

Training Details

Loss Function: CrossEntropyLoss Optimizer: Adam Learning Rate: 1e-3 (fine-tuning) Epochs: 15 Batch Size: 32

---

Training Results

Accuracy Plot

Loss Plot

---

Confusion Matrix

Model 0

Model 1

---

Inference Pipeline

Steps:

1. Load image
2. Resize to 224x224
3. Normalize
4. Forward pass through ResNet18
5. Apply Softmax
6. Get prediction (Argmax)

---