# 31st place solution

My solution is an ensemble of two high-scoring public notebooks:
[https://www.kaggle.com/code/tc0000/birdclef-starter-notebook](https://www.kaggle.com/code/tc0000/birdclef-starter-notebook)
(LB .654)
[https://www.kaggle.com/code/aikhmelnytskyy/birdclef24-pretraining-is-all-you-need-infer](https://www.kaggle.com/code/aikhmelnytskyy/birdclef24-pretraining-is-all-you-need-infer)
(LB .632)
All credit to their approaches, amazing work!
Since one of the main challenges in this competition was managing the shakeup and preventing overfitting on the public leaderboard score, an average ensemble, both solutions with equal weight, was used. I applied averaging the ensembled predictions over each audiofile, weighting with 75% the average, and 25% the individual row. The ensemble ran in around 1h.
Personally, besides domain adaptation, I think this competition showcases the value of ensembles for generalizing over a limited test set, which can have amazing applications in many practical fields. Finally, although I improved a bit on the public score, my final improvement over the notebook by
[@tc0000](https://www.kaggle.com/tc0000)
on private was minimal, showing the resilience of their approach to overfitting to public. Again, credits to the author(s) of this work!