Perform training on strictly simulated data and evaluate on real XRFI'd data

[jkerrigan]

We want to see if a more robust model can be trained that doesn't take into account the XRFI algorithm biases.

[jkerrigan]

FCNvsXRFI.pdf
Training on an entirely simulated dataset performs reasonably well and could be improved upon with a much more robust simulated dataset. Additionally the partially flagged channels between freq. channels 800 and 1000 look to be having difficulty predicting to the edges. We may want to attempt padding the input frequency windows but this would obviously lead to a performance hit.

[jkerrigan]

Example of prediction comparisons. Orange is True Positives, Red is False Positives, White is False Negatives. Simulated dataset has been improved to some degree by adding additional variation to the bandpass, higher amounts of RFI, and data augmentation. Input carved data has been padded to 68x68 to improve edge prediction.