Iesus Davila

Vision-language model (Qwen2-VL-2B fine-tuned with LoRA) that looks at a fashion product photo and answers in catalog language: color, article type, category, gender, usage, season and a one-line description. Raises exact-match accuracy from 52% (zero-shot) to 90% on held-out products, and includes an inference toolkit for product sheets, top-k confidence, listing audits and batch auto-tagging on a single T4 GPU.

PyTorch reimplementation of Apple's FastViT (ICCV 2023), trained from random initialization on 11 weather classes. Reaches 89.2% test accuracy (0.992 macro ROC-AUC), and structural reparameterization folds every multi-branch block into a single conv for 1.76x faster inference on GPU with identical outputs. Parameter counts of all 7 variants match the timm reference.

Explainability study that fine-tunes ResNet-50 on 11 weather classes (93.4% test accuracy) while recording Grad-CAM, Grad-CAM++ and Score-CAM after every epoch. All CAM methods are implemented from scratch with forward/backward hooks, and the project measures how attention stabilizes and sharpens during training and where the model looks when it fails.