Fine-tuned google/efficientnet-b0 for 7-class skin lesion classification on HAM10000. Part of the DermAI explainability project comparing CNN and Vision Transformer explanations.
Hyperparameter tuned: dropout = 0.2, drop_connect_rate = 0.4, augmentation = True, phase1 lr = 1e-3, phase2 lr = 8e-4, weight decay = 0.01, batch size = 16, warmup ratio = 0.1
Best validation Macro F1 = 0.7759