This is an implementation of the AttnGAN in PyTorch, with some experimental additions and changes.

Dataset

• Download the Caltech-UCSD Birds-200-2011 dataset and extract it to the root folder of the project.
• Download metadata (includes captions) and copy its contents to the dataset folder.

Experimenting

• To train a DAMSM model, use the python -m src.main train-damsm <EPOCHS> <NAME> [OPTIONS] command. EPOCHS sets the number of training epochs, NAME is the name the model is going to be saved with and further referenced by. Options include:

  • Set patience for early stopping: --patience=20
  • Set device: --device=cuda:0

• To train the GAN, use python -m src.main train-gan <EPOCHS> <NAME> <DAMSM> [OPTIONS]. EPOCHS and NAME are the number of training epochs and the name of the model respectively. DAMSM is the name of the DAMSM model to be used for text-encoding and auxiliary DAMSM-loss. Options include:

  • Continue training of a saved model: --gan=ExampleModelName
  • Set device: --device=cuda:1

• To generate an image for each sample in the test set, use python -m src.main validate-gan GAN DAMSM SAVEDIR [OPTIONS]. GAN and DAMSM are the names of the models to be used. SAVEDIR is the output directory. Options include:

  • Set device: --device=cuda:2

For different hyperparameters, change values in config.py.