SVHN Digit Classifier API

Overview

This project is a simple FastAPI-based web server that uses a pre-trained ResNet-18 model to classify handwritten or street-view house number digits from images. The model is trained on the SVHN (Street View House Numbers) dataset, which consists of digits 0-9. It provides an API endpoint to upload an image and receive the predicted digit.

Features

• FastAPI server for easy deployment and testing.
• Image upload and prediction using PyTorch and torchvision.
• Pre-trained model loaded from models/resnet18_svhn.pth.
• Supports RGB images resized to 32x32 pixels.
• Error handling for invalid inputs.

Project Structure

• server_model.py: Main FastAPI application with the prediction endpoint.
• models/resnet18_svhn.pth: Pre-trained ResNet-18 model weights for SVHN.
• models/cifar_net.pth: Additional model (not currently used).
• notebooks/notebook.ipynb: Jupyter notebook for model training or experimentation (currently minimal).

Requirements

• Python 3.8+
• FastAPI
• Uvicorn (for running the server)
• PyTorch
• Pillow (PIL)
• torchvision

Install dependencies:

pip install fastapi uvicorn torch torchvision pillow

Setup and Running

1. Clone or navigate to the project directory:

   cd D:\Projects\project
   

2. Ensure the model files are in the models/ directory. If not present, you'll need to train or download the resnet18_svhn.pth model.

3. Run the server:

   uvicorn server_model:app --reload --host 0.0.0.0 --port 8000

   • --reload: Enables auto-reload during development.
   • The server will be available at http://localhost:8000.

4. Test the API:

   • Visit http://localhost:8000/docs for interactive API documentation (Swagger UI).
   • Use the /predict POST endpoint to upload an image file.
   • Example curl request:

     curl -X POST "http://localhost:8000/predict" -H "accept: application/json" -H "Content-Type: multipart/form-data" -F "file=@path/to/your/image.jpg"

   • Response: {"prediction": "5"} (example).

Model Details

• Architecture: ResNet-18 with BasicBlock, adapted for 32x32 input images.
• Classes: Digits 0-9.
• Preprocessing: Resize to 32x32, normalize with mean/std (0.5, 0.5, 0.5).
• Inference: Runs on CPU (model loaded with map_location="cpu").

Training the Model

If you need to train the model yourself:

1. Use the notebooks/notebook.ipynb as a starting point.
2. Download the SVHN dataset using torchvision.
3. Train using standard PyTorch training loop with the ResNet architecture defined in server_model.py.

Troubleshooting

• If the model file is missing, the server will fail to load. Ensure models/resnet18_svhn.pth exists.
• For GPU support, modify the model loading to use CUDA if available.
• Image format: Supports JPEG, PNG, etc., converted to RGB.

License

This project is open-source. Feel free to use and modify.

---

Built with ❤️ using FastAPI and PyTorch