First download and install conda.

Create the conda environment by opening up a terminal in the /substorm-detection directory, then run

conda env create -f environment.yml

In pycharm, select the environment by going to File>Settings>Project>Project Interpreter>Project Interpreter: substorm-detection

• visualization
  • make "semantic dictionary"
    • see which neurons are being activated in an example
    • visualize those neurons with activation maximization
    • other stuff from distill
  • graph cnn won't have to deal with missing data so visualization might be easier
• regression
  • see what the model is choosing?
    • totally random
    • is this thrown off by outliers? would L1 loss help instead?
    • tried L1 Loss, didn't really help, pretty much no predictive power whatsoever, predictions in approximately the right range
  • why is it going so fast? compare with binary classification
    • make binary classification dataset
  • double check dataset
  • multiclass classification
  • initialize with binary classification weights
    • make binary classification dataset