CyCADA

This is unofficial implementation of CyCADA: Cycle-Consistent Adversarial Domain Adaptation (ICML2018).

Requirements

python >= 3.6
pytorch>= 1.0
torchvision

Setup dataset

I prepare a download code of MNIST->USPS dataset and run below.

python prepare_mnist2usps.py

If you conduct experiments on your dataset, please put data on the path: ../data/[your dataset] and specify dataroot option in scripts/train_cycada.sh (default: dataroot=../data/mnist_USPS)

Dataset structure must be

- [your dataset]
    - trainA
    - trainB
    - testA
    - testB

Domain A is source and domain B is target,
but if specifying direction="BtoA" in scripts/train_cycada.sh, switch source and target.

Directory structure

data: preprocess data and set loaders
options: set options for train and test phase
results: contain test results
util: pack useful functions
checkpoints: save training processes
models: model implementation

Pretraining

Pretrained models contain in pretrain
If you pretrain a source classifier before adaptation, please specify pretrain=1.

Train

If you conduct domain adaptation, please run below. All hyperparameters are packed.

./scripts/train_cycada.sh

Test

This test code automatically searches unevaluated models in checkpoints.

./scripts/test.sh

Model architecture

Generator: resnet-based networks with two residual blocks
Discriminator: 4-layers
Classifier: Revised LeNet for 32x32 images

Result

The result can be reproduced by using pretrained mnist, usps classifiers I set in pretrain.

models/pretrain/lenet_mnist_acc_97.5000.pt
models/pretrain/lenet_usps_acc_97.1599.pt

Model	Direction	M-U
Source-only	->	91.68
Source-only	<-	68.55
Cycada	->	96.0 (95.6)
Cycada	<-	95.0 (96.5)
Target-only	->	97.15
Target-only	<-	97.50

() denotes reference values in the cycada paper

Reference

Paper
CYCADA: CYCLE-CONSISTENT ADVERSARIAL DOMAIN ADAPTATION
In ICML, 2018
https://arxiv.org/pdf/1711.03213.pdf
Implementation
Code is mainly borrowed from junyanz/pytorch-CycleGAN-and-pix2pix
https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix

gboduljak / CyCADA