Côte d’Ivoire Byte-Sized Agriculture Challenge
Helping Côte d'Ivoire
7 000 EUR
Completed (10 months ago)
Classification
Object Detection
894 joined
163 active
Start
Apr 25, 25
Close
Jun 16, 25
Reveal
Jun 16, 25
Interesting challenge. Interesting data
Data · 22 May 2025, 08:21 · 5
This challenge is interesting for me in that I quickly built a model at the begining and got my highest score of 0.92x after a few iterations. As it was still experimental stuff, I didn't set any seed or do anything recommended for reproducibility.
I can no longer reproduce that score after setting seeds. In fact, my LB score is quite sensitive to the seed used - even if CV is somewhat stable. Has anyone else experienced the same?
0.92 in your first trial? Amazing. Given that We are limited to given test images and we can't download our own has made this challenge way tricky for me than when we had the choice to download the images ourselves. Using the given data I am getting very good cv scores but then there is like 0.10 cv gap, I don't know why? Wish we had the chance to download test images/data ourselves.
Is yours an cv or gbdt approach?
> Using the given data I am getting very good cv scores but then there is like 0.10 cv gap, I don't know why?
I also had this problem after setting seeds. But did some experimentation with data normalization to get the CV/LB gap from 0.1 to 0.02. Can reliably get 0.91x LB now.
> Is yours an cv or gbdt approach?
Using a CV approach so far.
I see, thanks.Goodluck!!
Great job @nymfree on reducing the CV/LB gap to 0.02!
I've been focusing on CV approaches these last few days. My split was based on crop placement and its target, but so far I haven’t found a clear correlation between CV and LB.
The LB seems really sensitive, which I guess is one of the drawbacks of F1 macro with a small/imbalance dataset.
Currently, I'm working with a single fold—my CV scores are around 93–94.8, but my LB varies from 85 to 88 for the same CV range.Is it the same for you?
For the normalisation, i used 2 well know strategies for satellite imagerie, minmax, and fixed.
i didin't notice a difference from my side...
Also, regarding your gap: was it achieved with a single fold, or using full cross-validation?
I think the gap was for a single fold.