NEXT: -> inadcl -> takes 600 iters to get off the ground!!
UPDATE flag_kookup.txt BECAUSE UNSUITABLE PHOTO MISSING!
-> then train with new additional data and rebalanced grad
-> misal -> 0 trainerr in 20 iter, 0.5 PCA, total imbalance stuck -> try again: -> fc7 only -> last time only 4 pos in val! make 15 -> imbalance: max redbox st 80% class, 50% blur -> try again: -> 70% class imb
-> inadcl 06 -> 0 trainerr in 20 iter, 0.5 PCA, total imbalance stuck -> try again: 97, 52, 74, 88 -> fc7 only -> imbalance: max redbox st 50% class, 70% blur -> maybe inadcl semantically hard though -> improve: -> imbalance: 68% class, 50% blur, graphic06
-> water 09 -> 0 trainerr in 20 iter, maybe 0.6 PCA but looks like noise -> try again: 79, 77, 78, 78 -> fc7 only -> imbalance: max redbox st no imbalance anywhere -> improve: -> max redbox 78% class imbalance -> fc7 then fc6? -> improve: NOT low risk
-
rebalanced gradient -> finetune from original ->
-
data augmentation: rotations -> easy 1st step: vertical mirror -> ON GRAPHIC07
-
ROC curve for optimal threshold -> modify run_classifier
DATA ISSUES: scrape ¬scrape | total ------+--------+---------|------- zone | 84,163 19,857 | 104,020 ¬zone | 16,480 6,497 | 22,977 ------+--------+---------| total |100,643 26,354 | 126,997
-> need: scrape ¬scrape | ------+--------+---------| zone | | ¬zone | N/A | ------+--------+---------|