lindagaw / adda-v2

adda-pytorch

Implementation of "Adversarial Discriminative Domain Adaptation"(https://arxiv.org/abs/1702.05464) on pytorch

dataset

mnist -> usps
usps -> mnist

command

$ python train_src.py --config "A->B config-file"

#training a model on source domain data
"A->B config-file" should be usps2mnist.yaml or mnist2usps.yaml
val reports the accuracy on target data

$ python main.py --config "A->B config-file"

#adapt the model trained on source data to fit target data with ADDA method
"A->B config-file" should be usps2mnist.yaml or mnist2usps.yaml

performance on testing set

Method	mnist -> usps	usps -> mnist
source only	84% (1560/1860)	78% (7820/10000)
adapted	92% (1709/1860)	91% (9074/10000)