Domain Adaptation via Prompt Learning

[Paper] [Code]

Domain adaption via prompt learning (DAPL), extends from CLIP and CoOp, offers a simple solution to the domain adaption problem. The prompt consists of three parts: domain-agnostic context, domain-specific context, and class label (token).

• The domain-agnostic context represents general task information and is shared among all images.
• The domain-specific context represents domain information and is shared within each domain. With domain-specific context modeling the discrepancy between domains, our model is able to discard domain alignment.
• The class label distinguishes different categories.

Our method is to tailor the powerful CLIP for UDA by designing trainable domain-agnostic, domain-specific and class prompt. By learning the representation of the prompt, our method actually learns a conditional propability distribution to deal with distribution shift. Hence, our method learns different decision boundaries for each domain. Moreover, we show that this allows disentanglement of semantic and domain representation with contrastive learning.

Performance of DAPrompt

We evaluate our method on three benchmarks: VisDA-2017, mini-DomainNet and Office-Home.

How to Install

Our code is built based on the source code of CoOp. So you need to install some dependent environments.

# install clip
pip install ftfy regex tqdm
pip install git+https://github.com/openai/CLIP.git

# clone dapl
git clone https://github.com/LeapLabTHU/DAPrompt.git

# install dassl
git clone https://github.com/KaiyangZhou/Dassl.pytorch.git
cd dassl
pip install -r requirements.txt
pip install .
cd ..

# you may download clip weights and modify the path to clip weights in clip file, or it could be downloaded automatically

You may follow the installation guide from CLIP and dassl.

Download Datasets

VisDA is a dataset from VisDA 2017 challenge. It contains two domains, i.e., 152397 synthetic images and 55388 real images.

Home page for VisDA dataset

Download the VisDA-classification dataset

Office-Home is a dataset comprised of 4 domains and 65 categories. It has a total of 15,588 images.

Home page for Office-Home dataset

Downliad the Office-Home dataset

How to Run

We provide the running scripts in scripts/. Make sure you change the path in DATA and run the commands under CoOp/scripts/.

Training

The commond is in the file DAPL/scripts/main.sh, which contains six input arguments:

• DATASET takes as input a dataset name, like visda or officehome. The valid names are the files' names in DAPL/configs/datasets/. The names of dataset, source domain and target domain is defined in the file. The visual backbone is also defined in the yaml. file for the common visual backbone is related to dataset. You may follow these files to establish new datasets;
• CFG means which config file to use, such as ep25-v0-32 (see DAPL/configs/trainers/DAPL/). The implemntation details are included in this file. You may modify the hyper-parameters in the file;
• T: is a temperature in the Eq. 1;
• TAU: is the pseudo label threshold in Eq. 9;
• U: is a coeffcient term of $\mathcal{L}_u(\mathcal{D}^d)$ in Eq. 10;
• NAME: defines the name of the task.

Below we provide examples on how to run DAPL on VisDA-2017. The file DAPL/scripts/main.sh defines the path to dataset in the line 6. You may set it as the true path to your dataset. If you want to train DAPL on the VisDA-2017 dataset, you may run the below command in the path DAPL/scripts:

bash main.sh visda17 ep25-32-csc 1.0 0.5 1.0 t0

Load a pre-trained Model

We have upload a pretrained weight. You can load it and evaluate in the target domain. The command is

bash eval.sh visda17 ep25-32-csc 1.0 0.5 1.0 t0

How to Develop New Algorithm

The structure of our lib:

• configs: the config file for datasets and trainers;
• dataset: the definition of datasets;
• scripts: bash command to train;
• trainers: the main file for training.

If you want to define a new method NewDA, you may need to develop the project according to the following guide:

• Implement the method in the DAPL/trainers/newda.py. The method should inherit the class TrianerXU (in the lib dassl/dassl/engine/trainer.py), and you may rewrite some functions: build_model, run_epoch, forward_backward, test. We recommand you to inherit the methods from TrianerXU;
• Import your method in the file DAPL/trainers/__init__.py;
• Import your method in the file DAPL/train.py;
• Make a new folder DAPL/trainers/NewDA/ and define the config file in the file DAPL/trainers/NewDA/ep25.yaml. You can name the yaml. file according to the hyper-parameters;
• Replace the "TRAINER" in the DAPL/scripts/main.sh with your new method(e.g. NewDA);
• You could run you code on VisDA-2017 by the command:

bash main.sh visda17 ep25 1.0 0.5 1.0 t0

If you want to add new dataset NewData, you may follow:

• Define a new dataset in the DAPL/datasets;
• Import your new dataset in the file DAPL/datasets/__init__.py and DAPL/train.py;
• Define the hyper-parameters of the dataset in the DAPL/configs/datasets/NewData.yaml;
• Replace the "DATA" in the DAPL/scripts/main.sh with your new dataset path(e.g. NewData).

For new algorithm development, there are some dependency useful to read:

• dassl/dassl/engine/trainer.py: defines the class TrainerXU;
• dassl/dassl/config/defaults.py: defines the default configs;
• dassl/dassl/data/data_manager.py and dassl/dassl/data/samplers.py: define the dataset and dataloader;
• dassl/dassl/engine: defines some da/dg/ssl methods.

Acknowledgement

Thanks for the following projects:

• CLIP
• Dassl
• CoOp

How to Contect Us

You can send an e-mail to gecj20 at mails.tsinghua.edu.cn if you have queations.

How to Cite DAPL

If you use this code in your research, please kindly cite the following paper

@article{ge2022domain,
  title={Domain Adaptation via Prompt Learning},
  author={Ge, Chunjiang and Huang, Rui and Xie, Mixue and Lai, Zihang and Song, Shiji and Li, Shuang and Huang, Gao},
  journal={arXiv preprint arXiv:2202.06687},
  year={2022}
}