How wongzbb/metateacher works
Project type and purpose
• This is a research‐oriented PyTorch implementation of “MetaTeacher,” a multi‐model domain‐adaptation pipeline for medical image classification.
• It trains one or more teacher networks (e.g., ResNet, DenseNet) and coordinates their knowledge to guide a smaller student network via specialized loss functions.
File structure
• Top‐level scripts:
• train.py – main training entry point (coordinated domain adaptation).
• train_source.py – source‐only pre‐training.
• data_list.py – dataset definitions and splits.
• loss.py – overall loss orchestrator.
• metric/loss.py – custom metric‐based losses (e.g., domain discrepancy).
• Core packages:
• models/ – definition of all model classes.
– embedding.py – shared feature embedder.
– student/ – small student network variants (lenet, mynet, fcnet, etc.).
– teacher/ – large teacher network variants (ResNet, VGG, AlexNet, etc.) and hint modules.
README for additional insights
• Confirms multi‐model “MetaTeacher” approach.
• Points to the NeurIPS paper for algorithmic details.