Cycle Consistent Adversarial Domain Adaptation (CyCADA)

A pytorch implementation of CyCADA.

If you use this code in your research please consider citing

       authors = {Judy Hoffman and Eric Tzeng and Taesung Park and Jun-Yan Zhu,
             and Phillip Isola and Kate Saenko and Alexei A. Efros and Trevor Darrell},
       title = {CyCADA: Cycle Consistent Adversarial Domain Adaptation},
       booktitle = {International Conference on Machine Learning (ICML)},
       year = 2018


  • Check out the repo (recursively will also checkout the CyCADA fork of the CycleGAN repo).
    git clone --recursive cycada
  • Install python requirements
    • pip install -r requirements.txt

Train image adaptation only (digits)

  • Image adaptation builds on the work on CycleGAN. The submodule in this repo is a fork which also includes the semantic consistency loss.
  • Pre-trained image results for digits may be downloaded here
  • Producing SVHN as MNIST
    • For an example of how to train image adaptation on SVHN->MNIST, see cyclegan/ From inside the cyclegan subfolder run
    • The snapshots will be stored in cyclegan/cycada_svhn2mnist_noIdentity. Inside set the epoch value to the epoch you wish to use and then run the script to generate 50 transformed images (to preview quickly) or run all to generate the full ~73K SVHN images as MNIST digits.
    • Results are stored inside cyclegan/results/cycada_svhn2mnist_noIdentity/train_75/images.
    • Note we use a dataset of mnist_svhn and for this experiment run in the reverse direction (BtoA), so the source (SVHN) images translated to look like MNIST digits will be stored as [label]_[imageId]_fake_B.png. Hence when images from this directory will be loaded later we will only images which match that naming convention.

Train feature adaptation only (digits)

  • The main script for feature adaptation can be found inside scripts/
  • Modify the data directory you which stores all digit datasets (or where they will be downloaded)

Train feature adaptation following image adaptation

  • Use the feature space adapt code with the data and models from image adaptation
  • For example: to train for the SVHN to MNIST shift, set src = 'svhn2mnist' and tgt = 'mnist' inside scripts/
  • Either download the relevant images above or run image space adaptation code and extract transferred images

Train Feature Adaptation for Semantic Segmentation

CyCADA pixel+feat SVHN2MNIST test(ckevin4747)


View Github