Cycle Consistent Adversarial Domain Adaptation (CyCADA)

A pytorch implementation of CyCADA.

If you use this code in your research please consider citing

@inproceedings{Hoffman_cycada2017,

       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

}

Setup

• Check out the repo (recursively will also checkout the CyCADA fork of the CycleGAN repo).

git clone --recursive https://github.com/jhoffman/cycada_release.git 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
  • SVHN as MNIST (114MB)
  • MNIST as USPS (6MB)
  • USPS as MNIST (3MB)
• Producing SVHN as MNIST
  • For an example of how to train image adaptation on SVHN->MNIST, see cyclegan/train_cycada.sh. From inside the cyclegan subfolder run train_cycada.sh.
  • The snapshots will be stored in cyclegan/cycada_svhn2mnist_noIdentity. Inside test_cycada.sh 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 test_cycada.sh 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/train_adda.py
• 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/train_adda.py
• Either download the relevant images above or run image space adaptation code and extract transferred images

Train Feature Adaptation for Semantic Segmentation

• Download GTA as CityScapes images (16GB).
• Download GTA DRN-26 model
• Download GTA as CityScapes DRN-26 model
• Adapt using scripts/train_fcn_adda.sh
  • Choose the desired src and tgt and datadir. Make sure to download the corresponding base model and data.
  • The final DRN-26 CyCADA model from GTA to CityScapes can be downloaded here