Work developed in Python using Jupyter Notebooks for the Neural Networks course at ITBA. The work consisted on a Kaggle challenge to classify images from the CIFAR-100 dataset.
The link to the Kaggle competition is https://www.kaggle.com/c/rn2021q1itba-cifar100.
For the competition, we tried several architectures, topologies or networks until we finally won using transfer learning and EfficientNetB3. The winner jupyter notebook is named test-09
In the analysis/ folder, there is a jupyter notebook were we analyzed the dataset to have a better understanding of the problem.
In the tests/ folder, there are other folders with jupyter notebooks for each attempt we did in order to reach a winner neural network.
- test-00: Base neural network with minor improvements
- test-01: TL with ResNet50 val_acc = 0.77
- test-02: TL with VGG16 not good results at first, we dropped this path
- test-03: Data augmentation (testing different libraries with the model created in test-00 and test-02)
- test-04: TL with EfficientNetB0 val_acc = 0.84
- test-05: Custom model with data augmentation val_acc = 0.63
- test-07: TL with EfficientNetB2 val_acc = 0.85
- test-09: TL with EfficientNetB3 val_acc = 0.87