Transfer Alignment Network for Blind Unsupervised Domain Adaptation

This package provides implementations of Transfer Alignment Network. Transfer Alignment Network is a stack of autoencoder, transfer aligner layers and mlp networks.

Code structure

./src/model: python scripts for model definition

./src/train: python scripts for train and test models defined in ./src/model

./src/demo: demo shell script for batch execution of training codes in ./src/train

Naming convention

auto_encoder (ae): Autoencoder

mlp: Multilayer Perceptron

v1: Multilayer Perceptron on top of Autoencoder

transfer aligner (aligner): Transfer Alignment Layer connecting source and target Autoencoder

v2: Multilayer Perceptron on top of Transfer Alignment Layer and Autoencoder

Usage

source ae_train -> source v1_train -> target ae_train -> target mn_train -> target v2_test

Dependencies

Numpy
TensorFlow

Data description

Download multivariate data from https://archive.ics.uci.edu/ml/datasets/

Output folder structure

Output of the model training is stored in the directory specified in argument log_dir
Output folder sturcture
- train.log: log file having results
- test.log: log file having results for every test step
- best.log: log file having only the best result
- hyperparameter: json file having hyperparameter configuration for the current step
- weight/: folder having csv files of trained weights
log file columns
- columns in log files for autoencoder and aligner training are loss, test loss, test_diff, test_rel_diff
- columns in log files for classifier training are loss, test loss, test_accuracy, auc_roc, auc_pr

Demo

There is a demo script src/demo/script.sh
- Input: data/{$data_name}/
- Output: src/results/step1, src/results/step2, src/results/step3, src/results/step4, src/results/test
  - log files for step1, 3, 4 are loss, test loss, test_diff, test_rel_diff
  - log files for step2 is loss, test loss, test_accuracy, auc_roc, auc_pr
  - log files for test is test loss, test_accuracy, auc_roc, auc_pr

Reference

If you use this code, please cite the following paper.

@article{XuK21,
  author    = {Huiwen Xu and U Kang},
  title     = {Transfer Alignment Network for Blind Unsupervised Domain Adaptation},
  journal   = {Knowl. Inf. Syst.},
  volume    = {63},
  number    = {11},
  pages     = {2861--2881},
  year      = {2021}
}

snudatalab/TAN