🐦 Bird or Not Classifier

A lightweight image classifier that tells you whether your image contains a bird, a forest, or neither!

Built using FastAI, Gradio, and Python, this app is ideal for a quick image check — powered by a fine-tuned computer vision model.

💡 How It Works

• The model was trained on images of birds and forests scraped using duckduckgo_search.
• I trained a resnet18 model using FastAI and exported only the model weights to a model.pth file (instead of model.pkl) for safe deployment on Hugging Face Spaces.
• The app uses a confidence threshold to return "Uncertain" if it's not sure the image is a bird or forest.

🖼 Try it Out

Upload an image and see the prediction:

• Bird image → "bird"
• Forest image → "forest"
• Other → "Uncertain – this might not be a bird or forest 🤔"

🧠 Model Info

• Trained on a small custom dataset of birds and forest images.
• Built with FastAI and exported with learn.export().
• Includes a confidence threshold — if the model isn't at least 85% confident, it returns "Uncertain".

📦 Installation

Clone the repo:

git clone https://github.com/Geff115/bird-or-not-classifier.git
cd bird-or-not-classifier

Create a virtual environment (optional but recommended)

python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows

Install dependencies:

pip install -r requirements.txt

🖼️ Usage

Launch the app locally:

python3 app.py

Visit http://127.0.0.1:7860 and upload an image to classify.

📁 Project Structure

bird-or-not-classifier/
├── app.py                # Gradio app
├── model.pkl             # Exported FastAI model
├── requirements.txt      # Dependencies
└── README.md             

🧪 Example Outputs

Image | Prediction

• 🐦 Bird (parrot) | bird
• 🌲 Forest Scene | forest
• 🚗 Random Car Photo | Uncertain – this might not be a bird or forest 🤔

🚀 Deployment Notes

I avoid using model.pkl due to unsafe pickle serialization in hosted environments like Hugging Face.
Instead, I manually reconstruct the model architecture in app.py and load model.pth.

🔮 Future Plans

• v2: Train with a third "other" class for better generalization.
• Add image augmentation + more diverse training data.
• Deploy to Hugging Face Spaces.

📄 License

MIT License — use freely for research, projects, or fun 🕊️