This is an implementation of Adversarial Discriminative Domain Adaptation https://arxiv.org/abs/1702.05464 using Keras for the purpose of visualizing the intermediate activations in a variational approximation of total correlation.
The class implementing ADDA is in adda.py.
usage: adda.py [-h] [-s SOURCE_WEIGHTS] [-e START_EPOCH]
[-n DISCRIMINATOR_EPOCHS] [-f]
[-a SOURCE_DISCRIMINATOR_WEIGHTS]
[-b TARGET_DISCRIMINATOR_WEIGHTS] [-t EVAL_SOURCE_CLASSIFIER]
[-d EVAL_TARGET_CLASSIFIER]
optional arguments:
-h, --help show this help message and exit
-s SOURCE_WEIGHTS, --source_weights SOURCE_WEIGHTS
Path to weights file to load source model for training
classification/adaptation
-e START_EPOCH, --start_epoch START_EPOCH
Epoch to begin training source model from
-n DISCRIMINATOR_EPOCHS, --discriminator_epochs DISCRIMINATOR_EPOCHS
Max number of steps to train discriminator
-f, --train_discriminator
Train discriminator model (if TRUE) vs Train source
classifier
-a SOURCE_DISCRIMINATOR_WEIGHTS, --source_discriminator_weights SOURCE_DISCRIMINATOR_WEIGHTS
Path to weights file to load source discriminator
-b TARGET_DISCRIMINATOR_WEIGHTS, --target_discriminator_weights TARGET_DISCRIMINATOR_WEIGHTS
Path to weights file to load target discriminator
-t EVAL_SOURCE_CLASSIFIER, --eval_source_classifier EVAL_SOURCE_CLASSIFIER
Path to source classifier model to test/evaluate
-d EVAL_TARGET_CLASSIFIER, --eval_target_classifier EVAL_TARGET_CLASSIFIER
Path to target discriminator model to test/evaluate
To run the source encoder model on MNIST, use:
python adda.py [-e START_EPOCH]
To run the discriminator model on MNIST and SVHN to increase domain confusion, use:
python adda.py -f [-s SOURCE_WEIGHTS] [-n DISCRIMINATOR_EPOCHS] [-a SOURCE_DISCRIMINATOR_WEIGHTS] [-b TARGET_DISCRIMINATOR_WEIGHTS]
python adda.py -t SOURCE_CLASSIFIER_WEIGHTS
python adda.py -t SOURCE_CLASSIFIER_WEIGHTS -d TARGET_DISCRIMINATOR_WEIGHTS