Deep learning under real-world constraints

PyTorch, transfer learning (DINOv3, BirdNET), LoRA, Optuna, GPU-side augmentation

Two PyTorch projects on messy, real-world data rather than clean benchmarks.

Petri-dish bacterial classification: a 16-class, small (4,692 image), imbalanced dataset. Compared frozen DINOv3 embeddings, DINOv3 with LoRA finetuning, and a from-scratch ResNet-18, reaching 71.68% hidden-test accuracy. A 4-factor augmentation search (120 Optuna trials) lifted the embeddings track from 0.63 to 0.72 accuracy. Honest limitation: the intended dish-orientation preprocessing step never worked.

BirdCLEF bioacoustic classification: 206 species, 35,549 recordings, severe class imbalance. The central result is a well-evidenced negative one - synthetic audio augmentation (AudioLDM2) improved neither of two independent model families, cross-validated with explicit leakage avoidance. Best model: BirdNET embeddings plus logistic regression, 0.8782 weighted F1.