Meeting Notes 2/6

[panalluri]

Diagnostics project:

• The dropoff for each filter should go almost vertically down after the filter cutoff <-- this is not happening
• Try the trial examples w GCM filters in the notebook link Dhruv sent --> apply what you learned to fix my graphs
• Use whiteboard pic to derive power spectra units w/ xrft
• Image in GCM filters documentation plots the filter kernel, not the filter cutoff like I thought
• Play with the GCM theory notebook in documentation --> plot the filter cutoff on a log scale
• air-sea flux anomalies --> don't use imshow
  • plot with lat-lon axes
  • the current plots are a mirror image of the earth
  • plot for some specific regions, not just the whole earth (e.g. gulf stream region, eastern pacific region off west coast of south america, 1000 x 1000 km)
  • narrow in on the colobar scale for freshwater flux (e.g. use plt.plot, robust=True)
  • make movies of the flux anomalies

Super-res project:

custom alpha loss:

• dig into every single line in CNN code + custom loss function w/ alpha and training loop and understand why it works
• minimum working example: try custom loss function alpha with a dummy MSE loss implementation + very simple data to try to debug (e.g. linear regression) --> this allows us to isolate the loss issue, and not get mixed into the CNN issues

fft loss:

• document what happens when only fft MSE is in loss func (no real space MSE)
• determine whether we need MSE for magnitude and phase or not

Overleaf:

• add board pics from last 3 meetings
• document results of each experiment

Paper:

• describe tianyi li et al paper to dhruv --> make a short ppt (45 mins ppt)
• read section A.3 to understand some conventions around spectra (https://www-sciencedirect-com.ezproxy.cul.columbia.edu/science/article/pii/S1463500317301221#sec0015)

Group meeting:

• 5-10 min presentation of where I am w my research