/Domain_adaptive_regression_GOL

Primary LanguagePythonApache License 2.0Apache-2.0

Domain adaptive regression with ordered embedding space

🤸 Visit our project page here

  • We propose to prepare the model for generalization by organizing the embedding space, using only source domain data.

  • The model is then adapted to unseen target domains with few-shot labels, using only target domain data.

  • We propose a generalization of manifold diffusion to regression tasks, which demonstrated some potential over simple distance-based measures like kNN, see below.

fig Figure 1: overview.

0. Dependencies

conda env create -f environment.yml

1. Procurement Stage

Train the feature extractor to establish ordered embedding space using source domain data:

python Procurement_train_GOL.py

For baseline comparison, train a vanilla regressor to predict values directly:

python Procurement_train_regressor.py

2. Deployment Stage

Few-shot adaptation on target domain data

  • No fine-tuning, no source domain access, 5*5=25 labels accessible in the target domain (main proposed setup in the paper)
python Deployment_few_shot_adaption.py
  • Alternatively, fine-tune the model with few-shot labels in the target domain
python Deployment_finetune_GOL.py

python Deployment_finetune_regressor.py

3. Manifold Diffusion for Regression

A distance measure in the high-dimensional manifold structure, taking into account both labeled and unlabeled data.

Code implementation in utils/diffusion.py:

class Diffuser(...)
    
    ...
    
    def diffuse_pred(...):
        ...

fig Figure 2: Manifold diffusion for regression vs. vanilla kNN label assignment.

Main code reference: