How greedylearner1146/hela-vfa works
Project Type and Purpose
• HELA-VFA is a PyTorch-based few-shot classification pipeline that implements a Hellinger-distance and attention-based variational feature aggregation network.
• It uses a ResNet-12 backbone with channel/spatial attention, a custom metric-learning loss library (the “Hesim” folder), and standard data-prep, augmentation, training, and evaluation scripts.
File-Structure Analysis
Top-level files & scripts:
• Data_Preparation.py, Data_Augmentation.py – image ingestion, splitting, transforms.
• Attention.py, ResNet12.py – attention modules and backbone definition.
• dataloader.py, sampler_loader.py, test_loader.py – few-shot episode construction.
• Hellinger_dist.py, HELA_VFA_main.py – computation of Hellinger distance and main variational aggregation.
• model_train.py, evaluate.py – orchestrate training loop and final evaluation.
• Hesim/… – metric-learning helper classes, reducers, distance functions culminating in HesimLoss.
README Insights
• Outlines a sequential pipeline: data prep → augmentation → attention model → backbone → sampling → loss → train → eval → test.
• External dependencies: PyTorch, Sicara few-shot library, PyTorch-metric-learning concepts.
• Data & weight files hosted on Google Drive.