Co-Teaching (CT)is a generic method for unsupervised domain adaptation and expansion. CT consists of knowledge distillation based CT (kdCT) and mixup based CT (miCT). Specifically, kdCT transfers knowledge from a leader-teacher network and an assistant-teacher network to a student network, to let the student better resolve cross-domain visual ambiguity, while miCT further enhances the generalization ability of the student. CT works for both image classification and image semantic segmentation.

Python version is 3.8.

pip install -r requirements.txt

First, we need download Office-Home and DomainNet into datasets folder. Our data division follows the KDDE. Download division data lists and pretrained checkpoints in this Google Drive link into VisualSearch.

We provide the command to inference model and predict one task.

python  predict.py  --config="configs/CT_DDC_ResNet50.yaml" --source="Art" --target="Clipart"  --dataset="officehome" --datasetroot="datasets/OfficeHome" --num_class=65 --run=1

We also provide the script to predict all tasks in one dataset (Office-Home or DomainNet).

bash predict.sh

We provide four different yaml files in configs，which can be used to configure the KDDE and CT. Changing different yaml files to train different models.

We provide the command to train one task.

python  train.py  --config="configs/CT_DDC_ResNet50.yaml" --source="Art" --target="Clipart"  --dataset="officehome" --datasetroot="datasets/OfficeHome" --num_class=65 --run=2

We also provide the script to train all tasts in one dataset (Office-Home or DomainNet).

bash train.sh

• ResNet50: Trained exclusively on the source domain.
• DDC: A classical deep domain adaptation model that minimizes domain discrepancy measured in light of first-order statistics of the deep features (Tzeng et al., Deep Domain Confusion: Maximizing for Domain Invariance, ArXiv 2014)
• SRDC: A deep domain adaptation model that enhances its discrimination ability by clustering features from intermediate layers of the network.
• KDDE: The first generic method for unsupervised domain adaptation and expansion.
• Co-Teaching: Our method for unsupervised domain adaptation and expansion by alleviating cross-domain visual ambiguity.

python eval_all_tasks.py --test_collection officehome_test

python eval_all_tasks.py --test_collection domainnet_test

This project is released under the Apache License. Please see the LICENSE file for more information.