Introduction

FoodVision Mini — an image classification project that recognizes food images as 🍕Pizza, 🥩Steak, or 🍣Sushi.

Instead of just experimenting in notebooks, I wanted to create something real:

• A modular deep learning pipeline
• A Gradio-powered interface
• A live deployment on Hugging Face Spaces

Whether you're just learning PyTorch or looking to deploy ML models, this blog walks you through the full-stack journey from training to deployment.

🔍 Project Overview

What does FoodVision Mini do?

Given an image, the model classifies it into one of three food categories using transfer learning with:

• EfficientNetB2 and
• Vision Transformer (ViT)

💡 I trained both models, evaluated their accuracy/speed, and deployed the better one for real-time use.

🏗️ Architecture

[Food101 Subset] --> [Transforms & Dataloaders] --> [EffNetB2/ViT] --> [Gradio Interface] --> [Hugging Face Deployment]

• Modularized with a src/ folder for reusable components
• CLI training scripts for reproducibility
• Clean separation of models, data, training, and UI

⚙️ Key Components

🧾 1. Dataset + Preprocessing

• Downloaded a small subset of the Food101 dataset (pizza, steak, sushi only)
• Applied resizing, normalization, and augmentation with torchvision.transforms

🧠 2. Model Training

I trained both:

• EfficientNetB2 using pretrained weights from torchvision
• ViT_B_16 for comparison on speed/accuracy

Each was trained via a dedicated CLI script:

1python scripts/train_effnetb2.py
2python scripts/train_vit.py

Results were saved and logged to .json files for visualization.

🧪 3. Evaluation

Plotted loss/accuracy curves and confusion matrices, and chose the better performing model based on:

• Inference time
• Validation accuracy

EffNetB2 Feature Extractor Loss/Accuracy Results

ViT Feature Extractor Loss/Accuracy Results

FoodVision Mini Inference: Speed VS Performance

🌐 4. Gradio UI

I built a modern Gradio interface:

• Upload or select example images
• Display top-3 predicted classes with confidence
• Prediction time visible for benchmarking

1demo = gr.Interface(
2    fn=lambda img: predict(img, model, transforms, CLASS_NAMES),
3    inputs=gr.Image(type="pil"),
4    outputs=[
5        gr.Label(num_top_classes=3, label="Predictions"),
6        gr.Number(label="Prediction time (s)"),
7    ],
8    examples=examples_list,
9    title=APP_TITLE,
10    description=APP_DESCRIPTION,
11    article=APP_ARTICLE,
12)

🚀 5. Deployment on Hugging Face Spaces

• Used git-lfs to manage .jpg and .pth files
• Enabled a smooth preview on spaces
• Auto-regenerates examples via generate_examples.py

🔥 Live Demo

🎯 Try the model in action:
👉 Live on Hugging Face
💻 GitHub Repo

💡 What I Learned

This project helped me grow in:

• 🔍 Computer Vision: Learned how transforms and dataset management affect performance
• 🔄 Transfer Learning: Understood how freezing layers vs. finetuning affects model output
• ⚙️ Modular ML Code: Clean separation of training, prediction, deployment logic
• 🧪 Deployment Best Practices: From Git LFS to Gradio interface testing on Spaces
• 🚀 Developer UX: Thought about UI feedback, image loading, and prediction clarity

📦 Key Skills

• PyTorch, TorchVision
• EfficientNet, ViT
• Gradio Interface/Blocks UI
• Git LFS + Hugging Face Deployment
• Modular ML Architecture

✅ Final Thoughts

FoodVision Mini taught me that a project doesn’t have to be huge to be powerful — what matters is delivering a full loop: from training to real-time usage.

This project serves as a starter template for anyone trying to build real-world deep learning pipelines with:

• Fast iteration
• Clean interfaces
• Reproducible results

Want to build one yourself? Start with just 3 classes — and scale up to Food101.

📎 Resources

• 📦 GitHub: https://github.com/RahulSaini02/pytorch-foodvision-101
• 🤖 Live Demo: https://huggingface.co/spaces/sainir/foodvision-mini
• 📸 Dataset: torchvision.datasets.Food101