Home

Awesome

chainer-ADDA

Implementation of Adversarial Discriminative Domain Adaptation in Chainer.

Note this code depends on this version of Chainer (or newer). Please check out the source from that link rather than installing via pip.

Results

The following results are for the SVHN to MNIST domain adaptation task. loss

Training% accuracy (Paper)% accuracy (This implementation)
source only0.6010.575
ADDA0.7600.800

Usage

Run python train.py -g 0 to train everything using GPU 0. SVHN and MNIST datasets will download automatically. If a classifier pretrained on the source (SVHN) domain is not found, one will be trained first, then continue on to do ADDA.

Resources