No description, website, or topics provided.
Branch: master
Clone or download
jhoffman Update usps.py
base url was missing so moved to other site where usps is stored
Latest commit 383f655 Oct 23, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
adda Update usps.py Oct 23, 2017
scripts Add script to run SVHN-MNIST experiment Mar 29, 2017
tools Add additional ADDA parameters and reorganize Mar 29, 2017
.gitignore Add missing data module Feb 24, 2017
readme.md Add basic readme Mar 29, 2017
requirements.txt Update requirements.txt Mar 27, 2017

readme.md

Adversarial Discriminative Domain Adaptation

Getting started

This code requires Python 3, and is implemented in Tensorflow.

Hopefully things should be fairly easy to run out of the box:

pip install -r requirements.txt
mkdir data snapshot
export PYTHONPATH="$PWD:$PYTHONPATH"
scripts/svhn-mnist.sh

The provided script does the following things:

  • Train a base LeNet model on SVHN (downloading SVHN under data/svhn in the process)
  • Use ADDA to adapt the SVHN model to MNIST (downloading MNIST under data/mnist in the process)
  • Run an evaluation on MNIST using the source-only model (stored at snapshot/lenet_svhn)
  • Run an evaluation on MNIST using the ADDA model (stored at snapshot/adda_lenet_svhn_mnist)

Areas of interest

  • Check scripts/svhn-mnist.sh for hyperparameters.
  • The LeNet model definition is in adda/models/lenet.py.
  • The model is annotated with data preprocessing info, which is used in the preprocessing function in adda/models/model.py.
  • The main ADDA logic happens in tools/train_adda.py.
  • The adversarial discriminator model definition is in adda/adversary.py.