oozdenizci / adversarialconditionalvae

Adversarially censored conditional variational autoencoders for invariant latent space learning.

Adversarially Censored Conditional Variational Autoencoders

This is an implementation of adversarially censored conditional variational autoencoder (cVAE) models for learning invariant latent spaces. Conventional cVAE model learning is performed within an adversarial training setting to monitor and/or censor nuisance-specific leakage in the learned latent space. Implementation is in Python using Keras with Tensorflow backend. It was implemented to be used with electroencephalographic (EEG) data from different subjects. This is a re-implementation of presented work (see below citation) with arbitrary encoder-decoder architectures.

Usage

An example execution is as follows:

from AdversarialcVAE import AdversarialcVAE

net = AdversarialcVAE(chans = ..., samples = ..., n_latent = ..., n_nuisance = ..., n_kernels = ..., adversarial = ..., lam = ...)

net.train(train_set, validation_set, log = ..., epochs = ..., batch_size = ...)

Parameter n_latent defines the latent space dimensionality, n_nuisance defines the nuisance variable dimensionality (i.e., the condition and the adversary output), and n_kernels define the number of convolutional kernel filters within the encoder and decoder which both have arbitrary structures for now. Boolean parameter adversarial = True trains the cVAE via adversarial censoring. If False, then an adjacent adversary network is simply trained to monitor nuisance-specific leakage in the latent representations. Parameter lam indicates the adversarial regularization weight embedded in the total loss function.

To use the train function, both train_set and validation_set should be three-element tuples (i.e., x_train, y_train, s_train = train_set). Here, the first element x_train is the EEG data of size (num_observations, num_channels, num_timesamples, 1), y_train are the one-hot encoded class labels (e.g., for binary labels will have a size (num_observations, 2)), and s_train are the one-hot encoded nuisance labels used for censoring (e.g., for 10-class nuisance labels will have size (num_observations, 10)). Variable log indicates the directory string to save the log files during training.

Paper Citation

If you use this code in your research and find it helpful, please cite the following paper:

Ozan Ozdenizci, Ye Wang, Toshiaki Koike-Akino, Deniz Erdogmus. "Transfer learning in brain-computer interfaces with adversarial variational autoencoders". 9th International IEEE/EMBS Conference on Neural Engineering (NER), 2019. https://dx.doi.org/10.1109/NER.2019.8716897