Skip to content

TropComplique/associative-domain-adaptation

Repository files navigation

This is a pytorch implementation of Associative Domain Adaptation.

Domain adaptation results

val ⬇️, train ➡️ just svhn just mnist svhn to mnist mnist to svhn
svhn 93.9 60.3 95.3 85.6
mnist 59.3 99.4 98.3 99.6

Notes

  1. I randomly color MNIST dataset during training and evaluation (see input_pipeline.py).
  2. To get "just svhn" and "just mnist" results run train_usual.py.
  3. To get "svhn to mnist" and "mnist to svhn" results run train.py.
  4. All four networks were trained for 15 epochs with the same hyperparameters.

Training curves

svhn to mnist mnist to svhn

Deep CORAL

Here are results for Deep CORAL: Correlation Alignment for Deep Domain Adaptation.

val ⬇️, train ➡️ svhn to mnist mnist to svhn
svhn 95.2 61.2
mnist 78.8 99.5

To get these results run train_deep_coral.py.

Requirements

  1. pytorch 1.0
  2. numpy, Pillow

Other implementations

  1. stes/torch-associative
  2. haeusser/learning_by_association