TimeVAE implementation in keras/tensorflow implementation of timevae:
TimeVAE is used for synthetic time-series data generation. See paper:
https://arxiv.org/abs/2111.08095
The methodology uses the Variational Autoencoder architecture. The decoder architecture is modified to include interpretable components of time-series, namely, level, trend, and seasonality.
'vae_conv_I_model.py' script contains the interpretable version of TimeVAE. See class 'VariationalAutoencoderConvInterpretable'.
'vae_conv_model.py' contains the base version of TimeVAE. See class 'VariationalAutoencoderConv'
The VariationalAutoencoderConvInterpretable in 'vae_conv_I_model.py' can also be used as base version by disabling the interpretability-related arguments during class initialization.
See script test_vae for usage of the TimeVAE model.
Note that 'vae_base' script contains an abstract super-class. It doesnt actually represent TimeVAE-Base.