This project focuses on building and evaluating Convolutional Neural Networks (CNNs) to classify images from the CIFAR-10 dataset. We experimented with optimizers (SGD and Adam), model depth, early stopping, and transfer learning with VGG16. A web app is also provided to test the final model.
- Baseline CNN with SGD (1-layer and deep): The shallow model underfits, while the deeper model improves performance.
- Deep CNN with Adam + EarlyStopping: Achieved the best performance (81.6% validation accuracy).
- Transfer Learning with VGG16: 73.27% accuracy using frozen layers.
- Testing: The final model was tested on 20 unseen images, correctly predicting 15/20.
| Model | Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|
| Deep CNN + Adam | 0.8076 | 0.8100 | 0.8076 | 0.8076 |
| Deep CNN + SGD | 0.7201 | 0.7346 | 0.7201 | 0.7174 |
| VGG16 (frozen) | 0.7327 | — | — | — |
File explanation:
project_1_deep_learning.ipynb - Our code where we build our model.
requirements.txt. - summary of libararies needed to run the code correctly.
Multiple model.h5 files - with our saved model results.
app.py - The app file that was used to build the app on HuggingFace spaces. See link below.
Report Project 1 - this word doc shows our report for this project.
Presentation - Our powerpoint slides for this project.
- Install dependencies:
pip install -r requirements.txt- Launch the notebook:
jupyter notebook project_1_deep_learning.ipynb- (Optional) Test the prediction app: 👉 https://huggingface.co/spaces/DaanBooy/Image_Predictor
Group 1: Daan, John-Bapist, Katy
This project is for educational purposes only.