Data augmentation:
- H, V, 90, 180, 270, T, T',
- scaling
- Feature Standardization - StdNormalize
- whitening
- more rotation
- shift translate - 15px doable Remove max and mix inputs in the angle encoding. I think we'll be back to before having the angle in performance. Concat angle as channel to image and feed to conv net.F Use RandomChoiceCompose to apply transpose Use RandomChoiceRotate istead of "range rotate" feed GMM Address class imbalance. Although test set may have a different balance
Replace NA with average of ships Try the reconstruction idea from Capsules paper Batchnorm layer settings? Remove to see if it helps Train on whole data before running on test data save model parameters
weird images: 3907 2691 8316 They seem to be translated/rotated Two peaks in histogram 1666 looks very much like an iceberg
dout: 0.1 [Epoch: 247 loss: 0.2920 | acc: 0.8704 | vloss: 0.3403 | vacc: 0.8625 ] dout: 0.2 [Epoch: 249 loss: 0.3447 | acc: 0.8465 | vloss: 0.3259 | vacc: 0.8656 ] dout: 0.2 - no rotation [Epoch: 249 loss: 0.2550 | acc: 0.8963 | vloss: 0.3291 | vacc: 0.8607 ] dout: 0.2 - no rotation 16-64 [Epoch: 249 loss: 0.2529 | acc: 0.8941 | vloss: 0.2576 | vacc: 0.9062 ] dout 0.2, 45deg, first and seconf conv are two, sigmoid after mean and std [Epoch: 249 loss: 0.3682 | acc: 0.8608 | vloss: 0.3038 | vadjloss: 0.2850 | vacc: 0.8656 | vacc_m: 0.8844 ]