Next go through modular folder edition should do these things!

[JustinGOSSES]

   - Find out why lots of wells are not being kept? Confirm it is no base pick by changing the initial feature creation main function to write reason for ignored well for each ignored well. 
   - Code to identify what tops and logs are common enough as separate notebook
   - Split train at start but after we know what wells we can use! Reasons to ignore wells include:
       - No top
       - No logs of the ones we want
       - Bad format, So can't load
   - More neighbor features, not just the thickness from nearest neighbor!
   - Run main feature creation function in backwards direction to get windows on other side of row (below in direction phrasing used in code)
   - Feature created by comparison of windows above and below
   - Add location based features?! May need map and geopandas draw to do this or other things?
   - Create additional normalised features with normalization at both well and all wells scales from features we already have!
   - Regroup training and test into easier to follow portions in ml part
   - Try umap feature reduction as separate path forward, also try PCA feature reduction
   - Currently dealing with wells that don't get any pick or near pick predictions by ignoring, maybe change this?

[JustinGOSSES]

EDIT:

1. Create new notebook for bullets 1,2,3 above.
2. Create slightly different notebook for nearest neighbor as described in bullet point 4.
3. Add additional features to feature creation notebook for points 5 and 6?
4. Redo machine-learning notebooks to make it easier to keep track of data sub-sets.
5. Try to understand what controls wells that don't get a classification prediction and reduce their numbers? Unsure what this will entail at this point in time.

[JustinGOSSES]

done most of these, others are in other issues