deepvats

deepvats copied to clipboard

Backend (Jupyter):

• Add a configuration parameter in the MVP notebook to choose among 3 options of including the matrix profile:
  1. Don't include it
  2. Include it in a way that only the matrix profile is used to train the encoder
  3. Include it along with the raw variables of the dataset

The matrix profile can be computed using the stumpy library. I think there's a way to compute it with GPU, you should look into it to accelerate the process.

Frontend: Ideally, for options 2 and 3 above, we would like to visualise somehow the computed matrix profile in the app...ideally, there should be a checkbox "show matrix profile" that plots the matrix profile instead of the time series data in the time series panel. To do that, the frontend should be able to access the computed matrix profile from the backend, which, as it is done with the rest of things, should happen through the storage module (i.e., through an artifact in wandb). This means that trains the encoder should not only log the encoder artifact, but also a new artifact corersponding to the computer matrix profile, so that the frontend can retrieve it and show it in the frontend.

For now, the frontend part is secondary (although very useful). You can still do the experiment implementing everything in the backend and using the app to compare the 3 modes mentioned above.