This is the official implementation of Self-supervised Image-to-text and Text-to-image Synthesis. The architecture of and are shown.

We use Caltech-UCSD Birds-200-2011 and Oxford-102 datasets in this work.

• Download Flower images
• Rename the jpg folder to images and unzip 102flowers.zip and put it inside 102flowers folder
• put 102flowers folder inside data folder
• Download Birds data and put inside Data/
• Download image data Extract them to Data/birds/

• pytorch
• torchvision
• tensorboardX
• pickle

The driver program for training the image autoencoder is main.py

python main.py --cfg cfg/flowers_3stages.yml --gpu 0

python main.py --cfg cfg/birds_3stages.yml --gpu 0

Models will automatically saved after a fixed number of iteration, to restart from a failed step edit netG_version in respective .yml file

python run_text_test.py dataset_type Input_Folder output_file.txt

• For Flower Dataset dataset_type=1, for Birds Dataset dataset_type=2 e.g.

python run_text_test.py 2 /home/user/dev/unsup/data_datasets/CUB_200_2011 outbirds_n.txt

python MappingImageText.py Dataset_folder

e.g.

python MappingImageText.py /home/user/dev/unsup/data_datasets/CUB_200_2011

python mmd_ganTI.py --dataset /home/das/dev/data_datasets/birds_dataset/CUB_200_2011 --gpu_device 0

python mmd_ganIT.py --dataset /home/das/dev/data_datasets/birds_dataset/CUB_200_2011 --gpu_device 0