DPStyler: Dynamic PromptStyler for Source-Free Domain Generalization

Datasets

• PACS: https://drive.google.com/uc?id=1m4X4fROCCXMO0lRLrr6Zz9Vb3974NWhE
• VLCS: http://www.mediafire.com/file/7yv132lgn1v267r/vlcs.tar.gz/file
• OfficeHome: https://drive.google.com/file/d/0B81rNlvomiwed0V1YUxQdC1uOTg/view?resourcekey=0-2SNWq0CDAuWOBRRBL7ZZsw
• DomainNet: http://ai.bu.edu/DomainNet/

Note: After downloading the data, unzip it into the 'datasets' folder.

Requirements

How to Start:

conda env create -f environment.yml

Training & Inference

Here, taking the training and inference of PACS as an example:

# Training:
cd scripts
sh run_clip_pacs.sh

Please note that the current default text template used for training is 'a CLS in a X style'. If you intend to perform multi-model ensemble inference, modify the text in 'DPStyler/dassl/txts/text_template.txt'.

Here are three text templates used in this paper:

1. a CLS in a X style
2. a X style of a CLS
3. a photo of a CLS with X like style

---

# Inference:
cd scripts
sh run_clip_pacs_test.sh

Model Weights

Model weights can be downloaded at https://drive.google.com/drive/folders/17C5NbTMGHRWkrH68nWFoxMVEOo1SCrRl?usp=share_link .