G-SFDA

Code (based on pytorch 1.3) for our ICCV 2021 paper 'Generalized Source-free Domain Adaptation'. [project] [paper].

Dataset preparing

Download the VisDA and Office-Home dataset. And denote the path of data list in the code.

Training

First train the model on source data with both source and target attention, then adapt the model to target domain in absence of source data. We use embedding layer to automatically produce the domain attention.

sh visda.sh (for VisDA)
sh office-home.sh (for Office-Home)

We provide the training log files, source model and target model on VisDA in this link. You can directly start the source-free adaptation from our source model to reproduce the results.

Domain Classifier

The file 'domain_classifier.ipynb' contains the code for training domain classifier and evaluating the model with estimated domain ID (on VisDA).