First of all, it depends on the split strategy you're using. Could you please clarify which one you’re following?
For us, it’s a split by location + target.
With this approach, We're currently able to reach the current LB score with 93% Accuracy and 93 F1 Macro CV Scores.
Just to confirm—did you reach the current LB score with only 88 CV? If so, I suspect there might be some overfitting to the public LB, unless you’re applying some sort of post-processing. Could you shed some light on that?
we can consintently get 0.96+ CV , based on a 5fold stratified KFold based on the target, but the LB keeps on fluctuating model by model by a gap of 0.01 to 0.05
@Abdourahamane_ You raised the bar a bit higher ! It will be a sleepless weekend for some of us if you confirm it's a modeling score (Not post processing) !
Hi @ASSAZZIN , I'm not using any postprocessing for my current score , I 'm just ensembling the predictions of three powerful models trained on the satellite imagery.
Hello @Belal_Emad,
First of all, it depends on the split strategy you're using. Could you please clarify which one you’re following? For us, it’s a split by location + target.
With this approach, We're currently able to reach the current LB score with 93% Accuracy and 93 F1 Macro CV Scores.
Just to confirm—did you reach the current LB score with only 88 CV? If so, I suspect there might be some overfitting to the public LB, unless you’re applying some sort of post-processing. Could you shed some light on that?
@marching_learning, @Koleshjr, @nymfree, @Abdourahamane_, @Gozie, @NogginHq, @Brainiac Would love to hear how things are going on your side in this competition!
Hello, ASSAZZIN
I use ensembling for submission while not in CV, thank you for sharing!
we can consintently get 0.96+ CV , based on a 5fold stratified KFold based on the target, but the LB keeps on fluctuating model by model by a gap of 0.01 to 0.05
@ASSAZZIN my CV is around 95 based on 10 folds stratified on the target.
Thanks @Belal_Emad, @Koleshjr, @Brainiac.
@Abdourahamane_ You raised the bar a bit higher ! It will be a sleepless weekend for some of us if you confirm it's a modeling score (Not post processing) !
Yeah that's an impressive score for sure .
Hi @ASSAZZIN , I'm not using any postprocessing for my current score , I 'm just ensembling the predictions of three powerful models trained on the satellite imagery.
My CV score is almost perfect.
@Abdourahamane_ Bravo !
Very impressive!
wooow impressive