TODOS

[ramarty]

General TODOS:

• Follow Yeh's code for getting TFrecords // adapt for sentinel
• Standardize NTL bands
• OSM - Extract data (check other paper. try angles? -- (1) cardinal directions, (2) absolute value equivalent; 90 = 270, (3) standard deviation... but up and sideways may still have high SD?). Could do how compares to larger/surrounding area.
• Get Keras working for individual npys [LINK]. Multiple inputs (stack link); more useful stack here
• For prepping CNN, add skip if already processed
• Transfer learning dataset -- beyond just survey locations
• Can use use larger than 224 for VGG16? If not, what to do about sentinel (survey 4 locations?). Also some sentinel bands aren't 10m
• Facebook data - if didn't scrape all params, delete
• DHS: globally consistent asset index (check what other paper does - Yeh et al)
• Add CNN for NTL
• Check RGB npy - visualize, should look as expected

ML Model TODOS:

• Grab actual values from tfrecords (so can compare with predicted)
• Standardize tfrecord values

[ramarty]

Include all bands, including NTL, in one numpy array. Can then grab each one during data generation! Also include wealth vars? Hmmm not needed if named. Put each in country-year folder.