The code in this repository is designed to work with datasets: Airbnb and PetFinder. The embedding-images before transfer is transformed by create_img.sh and are saved in each dataset folder.
The code in this repository is designed to work with Baselines: SDEdit and DDRM. We pretrained score-based generative models using the DDIM framework for each modality and both baseline models leverage the same pretrained model. The embedding-images after transfer are saved in the trans folder.
Both baselines shared the same downstream model, which are saved in the classifiers folder. Inference results are saved in the logs&result. Used train_clf.sh and inference.sh to train downstream model and inference.
Visualization of embeddings transfer using UMAP for several combinations are saved in visualization folder. Used visualize.sh to visualize other modalities transfer.
See the hyperparameters folder for baseline configs and args settings.