yes i got this score without using 'damage' column neither in training nor in testing.

Greate score! What method did you use? How long is your training?

i treated this probem as binary classification by dividing targets by 100 . training time is around ~3 hours for single model on p100 gpu and final score is ensemble of 2 models .

Nice one. This is a very clever way of solving this. Thanks for sharing

Congratulations! I think you will get a top-5 score after all disqualifications.

What do you mean by dividing a target by 100? What will be the class for value 0.9, 0, or 1?

I briefly tried an approach of training a classifier for `extent` 0 vs > 0 and I'm not sure if your approach was the same or not.

I had a similar approach. It's motivated by a few Kaggle notebooks which recommended this - you set target as (extent / 100), and train with binarycrossentropy. so the "class value" is the extent

Great solution @Agastya and yes I believe you'll get a top 5 placement on the LB once zindi clarifies issues with the damage feature (their previous explanation does little to iron out issues as previously raised and also considering the goal of the competition).

I personally have two subs one completely not relying on the damage feature and the other using pre and post processing with the damage feature, so I expect a drop once they provide a clearer explanation on what they'll consider a leak taking into account the goal of the competition.

I don't understand your solution very well. Can you explain more clearly? And importantly, did you use data leaks in your solution?

I did not use dataleak. I think this can be approached by an ensemble of 2 CNNs. You basically create a new column on your df to indicate whether an image has an "extent" value of 0 or not. You use this column as target to train the first CNN for binary classification which classifies an image as having an extent of 0 or 1. The output will be used as input to train the second model to predict the value of the non-zero images. Now, this is just something I though of, and not what I used for the competition. My own model did not perform very well, which was just a normal CNN.

Oh thank you! So your solution was just simple a CNN with regression output?

I tried this too but couldn't get this to work. nice job!

@Ahmadmwali I see you said your score is about 13, but on the LB I see it is 9.27/9.98. I'm not sure what the score of 13 is that you're talking about? Thanks

Just checked the leaderboard. So I saw that they selected the first model I made, which had data leak. My later models were not trained with it, which gave me about 13 RMSE. I forgot to select the later submissions. I'm getting disqualified as well.

@Ahmadmwali Oh thank you! I have learned a lot from your new idea in CNN (classification + regression). I appriciated it.