earhian/Humpback-Whale-Identification-1st-

Employs a multi-stage training pipeline with SENet154 backbone, progressively unfreezing layers and reducing learning rates to optimize metric learning for fine-grained whale identification. Integrates mask-based data preprocessing and ensemble predictions via CSV outputs, leveraging PyTorch 0.4.1 with supplementary playground dataset augmentation to improve generalization across whale flukes.

640 stars. No commits in the last 6 months.