Deep learning on raw signal data per cycle

[joepvdbogaert]

This PR adds the following functionality, mostly aimed at applying deep learning on the raw signal:

• engineering module: added functions to find the earthquakes in the training data and save the 'cycles' in between earthquakes as separate pickle files for quick reading.
• deep module: a new module with two helper classes and two methods to perform training and evaluation on the cycles:
  • Scaler: scales the data in various ways, which is crucial for DL to work. By varying the scaling method, we can easily experiment with different techniques.
  • KFoldCycles: splits the cycles in train and validation cycles over multiple folds to perform cross validation without data leakage.
  • train_on_cycles: trains a model on specified cycles (e.g., those suggested by KFoldCycles.split()).
  • evaluate_on_cycles: evaluates a model in the same way.

Finally, I added a notebook DeepEarthquakes.ipynb that shows how exactly to use the new code.

[san89]

Well done, nice written and efficient code.

My only comment is that I think that making an effort to centralize certain functionalities is worth it.