🐶 Dog Breed Identifier

This project demonstrates how to build a robust Transfer Learning pipeline using ResNet50 for classifying 120 dog breeds from the Stanford Dogs Dataset. The final model achieves ~81% accuracy on the validation set using only ~9,600 training images. The trained model is then deployed via a lightweight Streamlit App to perform custom image inference.

📌 Project Highlights

• Model Architecture: Pretrained ResNet50 + custom classifier head with ReLU and Dropout
• Dataset: 120 dog breeds (~12K training images)
• Training Accuracy: ~82.15%
• Validation Accuracy: ~80.51% at Epoch 29/30
• Loss Curve: Smooth convergence without overfitting
• Inference Interface: Web-powered Streamlit App
• Prediction Output: Predicted dog breed + Confidence score
• Visual Feedback: ✅ or ✳️☑️ label color based on confidence threshold

🧠 Model Training Summary

• Base model: ResNet50 pretrained on ImageNet
• Freezing: All ResNet convolutional layers frozen
• Classifier head:

  def setup_resnet50_for_transfer_learning(num_classes: int):
    model = models.resnet50(pretrained=True)
    
    # Freeze all parameters in the pre-trained layers
    for param in model.parameters():
        param.requires_grad = False
  
    # Replace the final fully connected layer (classifier)
    # The 'fc' layer in ResNet50 is the final classification layer
    num_ftrs = model.fc.in_features
    model.fc = nn.Linear(num_ftrs, num_classes)
    return model

Getting Started

pip install -r requirements.txt

Project structure

. ├── Appv2.py # Streamlit App ├── resnet50TL.ipynb # Model training notebook ├── requirements.txt # Dependencies ├── resnet50_best.pth # Saved model ├── data/ │ └── train/ # 120 folders, each named e.g. 'n02110185-siberian_husky' │ └── test/ ├── custom_images/images/ # Upload folder for app inference

Running the App

streamlit run Appv2.py

Sample Inference

• Uploaded Image: golden.jpg
• Prediction: golden_retriever
• Confidence: ✅ A figure

Future Improvements

• Fine-tune top layers of ResNet
• Evaluate ensemble or lightweight models (e.g., EfficientNet-B0)

Dockerize for deployment

Acknowledgements

• Stanford Dogs Dataset
• PyTorch
• Streamlit
• torchvision.models

License

• This project is licensed under the MIT License — see LICENSE file for details.