AlexNet on 102 Category Flower Dataset

Code includes whole assignment and process of hypertuning the AlexNet in comparison to basic convolutional network.

Oxford 102 Flower Dataset

What was the process?

• Testing basic CNN's
• Testing baseline AlexNet
• SGD vs Adam optimizer comparison
• Simplified hyperparameters grid search with early stopping mechanism
• Data augmentation

Weights and biases reports

• Report 1: Learning rates
• Report 2: Dropout rates
• Report 3: Weight decay
• Report 4: Data augmentation

Final results on test dataset

• AlexNet: 51.97% accuracy (manually stopped after 100 epochs)
• Simple CNN: 20.81% accuracy