We recommend to use conda environment. Please run

conda env create --file IIAE.yml

Download MNIST-CDCB, Facades, and Maps datasets.

Download Sketchy Extended dataset by following SEM-PCYC.

Lastly, download Cars dataset and VGG features of Skethcy Extended dataset from this link.

Note that we extracted VGG features from the pretrained model provided by SEM-PCYC.

Once you download all those datasets, please make sure you locate them under <root_path>/dataset/.

EX) <root_path>/dataset/MNIST-CDCB

​ <root_path>/dataset/cars

​ <root_path>/dataset/Sketchy

​ <root_path>/dataset/SketchyVGG

Now you are ready to train Interaction Information AutoEncoders (IIAEs).

You can also find the pretrained IIAEs for those five different datasets from this link.

Train IIAE with Skethcy dataset.

python zs_cross_vae.py --mode train --root_path <root_path> --output_dir <root_path>/SUBMISSION/Sketchy

Evaluate IIAE with Skethcy dataset.

python zs_cross_vae.py --mode test --root_path <root_path> --output_dir <root_path>/SUBMISSION/Sketchy

Train IIAE with any of MNIST-CDCB, Cars, Facades, or Maps.

python run_cross_vae.py --mode train --dataset_name <dataset_name> --root_path <root_path> --output_dir <root_path>/SUBMISSION/<dataset_name>

Extract features and evaluate cross-domain retrieval (for MNIST-CDCB, Facades, and Maps).

python run_cross_vae.py --mode features --dataset_name <dataset_name> --root_path <root_path> --output_dir <root_path>/SUBMISSION/<dataset_name>
python feat_dist_full.py --dataset_name <dataset_name> --root_path <root_path> --output_dir <root_path>/SUBMISSION/<dataset_name>

Generate samples of cross-domain image translation. (for MNIST-CDCB and Cars)

python run_cross_vae.py --mode test --dataset_name <dataset_name> --root_path <root_path> --output_dir <root_path>/SUBMISSION/<dataset_name>

Our code is based on Cross-domain Disentantanglement Network.

The VGG features of Sketchy dataset are extracted using pretrained VGG model provided by SEM-PCYC.

Raw files of Cars dataset can be found here.