Sijuade/imagenet-resnet50-inference

ImageNet Vision AI 🖼️

A beautiful, production-ready Streamlit application for ImageNet image classification using a custom-trained ResNet50 model achieving 76.21% top-1 accuracy.

✨ Live Demo

Try it now: ImageNet Vision AI on Hugging Face Spaces

🚀 Features

🎨 Beautiful User Interface

• —Modern Design: Purple gradient theme with smooth animations
• —Drag & Drop: Intuitive interface for image uploads
• —Multiple Images: Upload and process multiple images simultaneously
• —Prediction Cards: Stunning card-based design with medals (🥇🥈🥉) for top predictions
• —Color-Coded Confidence: Visual indicators (green/orange/gray) based on confidence levels
• —Sample Gallery: 4 pre-loaded ImageNet sample images for quick testing

🔥 Core Functionality

• —Top-K Predictions: View 1-10 top predictions per image
• —Confidence Filtering: Filter predictions by confidence threshold
• —Inference Metrics: Display inference time for each image
• —Batch Processing: Process multiple images in one session
• —Batch Export: Download all predictions as JSON

🎯 Model Details

• —Architecture: ResNet50 (PyTorch)
• —Training: Custom training on full ImageNet-1K dataset
• —Accuracy: 76.21% top-1 accuracy
• —Classes: 1000 ImageNet categories
• —Framework: PyTorch Lightning with mixed precision (FP16)
• —Optimizer: LAMB with cosine learning rate schedule
• —Augmentation: RandAugment, Mixup (0.2), CutMix (1.0), Random Erasing (0.25)
• —Training Details: 300 epochs, 8 GPUs (DDP), batch size 256

📦 Model Repository

• —Checkpoint: Sijuade/resnett50-imagenet
• —Training Code: ImageNet-Full-training

🎨 How to Use

1. 1.Upload Images: Drag and drop images or click to browse (JPG, JPEG, PNG, WebP)
2. 2.Try Samples: Click on any of the 4 sample images to see instant predictions
3. 3.Adjust Settings: Use sidebar to configure:
4. 4.Top-K predictions (1-10)
5. 5.Confidence threshold (0-100%)
6. 6.Show/hide inference time
7. 7.Show/hide model info
8. 8.View Results: See beautiful prediction cards with:
9. 9.Class names
10. 10.Confidence scores
11. 11.Visual progress bars
12. 12.Medal indicators for top 3
13. 13.Export: Download all predictions as JSON

🛠️ Technology Stack

• —Frontend: Streamlit with custom CSS
• —Backend: PyTorch + torchvision
• —Model: ResNet50 (PyTorch Lightning checkpoint)
• —Deployment: Docker on Hugging Face Spaces
• —Sample Images: From imagenet-sample-images

🐳 Docker Deployment

This app is deployed using Docker on Hugging Face Spaces. Key features:

• —Automatic Model Download: Downloads 391MB checkpoint from Hugging Face Hub during build
• —Sample Images: Downloads 10 sample images from GitHub during build (limited to 4 in UI)
• —User Permissions: Runs as user ID 1000 (HF Spaces requirement)
• —Port: Exposes port 7860 (HF Spaces default)

Dockerfile Highlights

# Download model from HF Hub
RUN python -c "from huggingface_hub import hf_hub_download; \
    hf_hub_download(repo_id='Sijuade/resnett50-imagenet', \
                    filename='acc1=76.2100.ckpt', \
                    local_dir='/app/models')"

# Download sample images from GitHub
RUN python -c "import os, requests; \
    samples = ['n01440764_tench.JPEG', 'n01443537_goldfish.JPEG', ...]; \
    [open(os.path.join('/app/images', f), 'wb').write(requests.get(base_url + f).content) for f in samples]"

📊 Example Predictions

The model can classify:

• —🐕 Animals: Dogs, cats, birds, fish, reptiles, insects
• —🚗 Vehicles: Cars, planes, boats, trains, bicycles
• —🪑 Objects: Furniture, electronics, tools, instruments
• —🍎 Food: Fruits, vegetables, dishes, beverages
• —🌸 Nature: Flowers, plants, landscapes, natural formations

🔧 Local Development

Using Docker

# Clone the repository
git clone https://github.com/cydal/imagenet-streamlit.git
cd imagenet-streamlit

# Build and run
docker build -t imagenet-vision-ai .
docker run -p 7860:7860 imagenet-vision-ai

Using Python

# Install dependencies
pip install -r requirements.txt

# Run the app
streamlit run app.py

📝 Training Your Own Model

Want to train your own ImageNet model? Check out the training repository:

[ImageNet-Full-training](https://github.com/cydal/ImageNet-Full-training)

Features:

• —Full ImageNet-1K training pipeline
• —PyTorch Lightning implementation
• —Multi-GPU (DDP) support
• —Mixed precision (FP16) training
• —Advanced augmentation techniques
• —Cosine learning rate scheduling
• —LAMB optimizer
• —Model compilation with torch.compile()

🔗 Links

• —Live App: Hugging Face Space
• —Model: Sijuade/resnett50-imagenet
• —Training Code: ImageNet-Full-training
• —App Code: imagenet-streamlit

📄 License

MIT License - See LICENSE file for details

🙏 Acknowledgments

• —ImageNet Dataset: ImageNet Large Scale Visual Recognition Challenge
• —PyTorch: Deep learning framework
• —Streamlit: Web application framework
• —Hugging Face: Model hosting and Spaces platform
• —Sample Images: EliSchwartz/imagenet-sample-images

Built with ❤️ using Streamlit & PyTorch | 76.21% Top-1 Accuracy on ImageNet-1K