This was a project for one of the modules on my masters degree. The Cifar dataset is a classic academic set to work on for image classification.
Jupyter notebook can be found here
My report can be found here
- Tensorflow
- Keras
- Sklearn
- Seaborn
- Numpy
- Pandas
- Matplotlib
I was able to test various Keras modules and tune hyperparameters in CNN models and data augmentation and transfer learning. I managed to get around 88 percent accuracy just by optimizing number of layers, max pooling and batch normalisation.
Training computer vision models takes time and resources! Initially I used the GPU on my Mac M1 but found it to be too slow, so I moved the trianing process to Google Colab and paid for some extra GPU power, and even then it took hours!
I also learned that sometimes the simple models work better and overcomplicating things does just that.
If you have any feedback, please reach out to me at mark@markstent.co.za