Official Implementation of PCT

Prerequisites

• python == 3.8.5

Please make sure you have the following libraries installed:

• numpy
• torch>=1.4.0
• torchvision>=0.5.0

Datasets

• Office-31
• Office-Home
• DomainNet

We provide direct download links in the script. However, for file larger than 100 MB (OfficeHome - Art and RealWorld), please manually download them from the following links--Art and RealWorld--and extract them appropriately.

Usage

• beta - learning rate/ momentum parameter to learn proportions in the target domain ( beta=0 corresponds to using a uniform prior)
• sub_s - subsample the source dataset
• sub_t - subsample the target dataset

Below, we provide example commands to run our method.

# Train PCT on Office-31 Amazon -> Webcam task using ResNet 50.
# Assume you have put the datasets under the path `data/office-31

# Single-source adaptation 
python examples/proto.py data/office31 -d Office31 -s A -t W -a resnet50  --epochs 10

# Sub-sampled source adaptation (uniform prior)
python examples/proto.py data/office31 -d Office31 -s A -t W -a resnet50  --epochs 10 --sub_s

# Sub-sampled source adaptation (learnable prior)
python examples/proto.py data/office31 -d Office31 -s A -t W -a resnet50  --epochs 10 --sub_s --beta 0.001

# Sub-sampled target adaptation (uniform prior)
python examples/proto.py data/office31 -d Office31 -s A -t W -a resnet50  --epochs 10 --sub_t

# Sub-sampled target adaptation (learnable prior)
python examples/proto.py data/office31 -d Office31 -s A -t W -a resnet50  --epochs 10 --sub_t --beta 0.001

Example commands are included in examples/proto.sh.

For source-private adaptation, please follow the instruction in the readme.md in the `Proto_Private' folder.

Citation

We adapt our code base from the v0.1 of the DALIB library.

If you find our framework useful, please cite our paper.

PCT

@inproceedings{tanwisuth2021prototype,
title={A Prototype-Oriented Framework for Unsupervised Domain Adaptation},
author={Korawat Tanwisuth and Xinjie Fan and Huangjie Zheng and Shujian Zhang and Hao Zhang and Bo Chen and Mingyuan Zhou},
booktitle = {NeurIPS 2021: Neural Information Processing Systems},
month={Dec.},
Note = {(the first three authors contributed equally)},
year = {2021}
}

DALIB

@misc{dalib,
author = {Junguang Jiang, Baixu Chen, Bo Fu, Mingsheng Long},
title = {Transfer-Learning-library},
year = {2020},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/thuml/Transfer-Learning-Library}},
}