This is just a repo to play around with VAEs. In this case I'm curious if a VAE is able to find things like phase shifts, amplitudes and frequencies in a couple of sin wave-like signals.

We have two time series. Both sin signals ($s_1 = a_1sin(2\pi f)$ and $s_2 = a_2sin(2\pi f + \phi)$). To generate the data we assume $a_1, a_2, f, and \phi$ to be uniform random variables in some interval. Now we use a BetaVAE and try to encode these four hidden states into the latend space.

Clone this repo, go to the project root and run (I'm assuming you run some unix system are using conda):

conda create -n vae-ts python=3.9
conda activate vae-ts
pip install -e . 

You'll first need to create the data set using this notebook: ./notebooks/01_datagen.ipynb. Now you can go a head an start a training job like this:

ipython vae_ts_test.vae.py

Note that the hyper parameters and some other constants are stored in the file ./constants.py. Here you might also adjust if you want to use GPUs and stuff.