pytorch 1.0 implementation for "augmented cyclic adversarial learning for low resource domain adaptation"

This repository is unofficial partial implementation of "augmented cyclic adversarial learning for low resource domain adaptation"(ICLR 2019, Ehsan Hosseini-As et al.)
We implemented the case of supervised setting, especially few shot setting.

• pytorch 1.0
• CUDA 10.0
• wandb

run

mkdir dataset
mkdir result
python train.py ./config/digit/config.yaml   

We treat few shot adaptation setting from SVHN to MNIST.
By exploiting wandb sweeping tool, we recoreded 80.5% accuracy on target domain while the paper reported to be about 84%. When 3 samples per class are given in target domain.