/Lifelong-Gan

A tensorlayer implementation of LifeLong Gan

Primary LanguagePythonMIT LicenseMIT

LifeLong-Gan

Requirements

  • Tensorflow 2.0.0
  • Tensorlayer 2.0
  • Python 3.6
  • Numpy
  • Tqdm

Model

Prepare data

  • Create a folder data in the project directory.
  • You can run download_dataset.sh in folder datasets to download the images or directly put the images in folder data, training data in a subdirectory train and test data in val.

Run

  • Training

    python train.py --tasks edges2shoes+facades

    Tasks edges2shoes and facades will be trained in turn. You can replace them with your datasets and concatenate the task names with "+". More training settings can be found in params.py.

  • Evaluation

    python evaluate.py --tasks edges2shoes+facades

    The evaluation results can be found in subdirectory samples.

Reference

  • [1] Lifelong GAN: Continual Learning for Conditional Image Generation. ICCV, 2019

  • [2] Toward Multimodal Image-to-Image Translation. NeurIPS, 2017.