My solution is an ensemble of two high-scoring public notebooks:
\nhttps://www.kaggle.com/code/tc0000/birdclef-starter-notebook (LB .654)
\nhttps://www.kaggle.com/code/aikhmelnytskyy/birdclef24-pretraining-is-all-you-need-infer (LB .632)
All credit to their approaches, amazing work!
\nSince 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.
\nPersonally, 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 on private was minimal, showing the resilience of their approach to overfitting to public. Again, credits to the author(s) of this work!
", "messages": [ { "id": "2866265", "postDate": "06/11/2024 08:25:08", "content": "My solution is an ensemble of two high-scoring public notebooks:
\nhttps://www.kaggle.com/code/tc0000/birdclef-starter-notebook (LB .654)
\nhttps://www.kaggle.com/code/aikhmelnytskyy/birdclef24-pretraining-is-all-you-need-infer (LB .632)
All credit to their approaches, amazing work!
\nSince 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.
\nPersonally, 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 on private was minimal, showing the resilience of their approach to overfitting to public. Again, credits to the author(s) of this work!
", "rawMarkdown": "My solution is an ensemble of two high-scoring public notebooks:\n\nhttps://www.kaggle.com/code/tc0000/birdclef-starter-notebook (LB .654)\nhttps://www.kaggle.com/code/aikhmelnytskyy/birdclef24-pretraining-is-all-you-need-infer (LB .632)\n\nAll credit to their approaches, amazing work! \n\nSince 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. \n\nPersonally, 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 on private was minimal, showing the resilience of their approach to overfitting to public. Again, credits to the author(s) of this work!", "votes": null }, { "id": "2866302", "postDate": "06/11/2024 08:39:04", "content": "Really good job! Upvoted!
", "rawMarkdown": "Really good job! Upvoted!", "votes": null }, { "id": "2866646", "postDate": "06/11/2024 12:38:23", "content": "Congratulations Sietse on your first silver medal!
", "rawMarkdown": "Congratulations Sietse on your first silver medal!", "votes": null } ], "comments": [ { "id": 2866302, "author_name": "mrsimple07", "author_url": "", "post_date": "06/11/2024 08:39:04", "content": "Really good job! Upvoted!
", "votes": null, "replies": [] }, { "id": 2866646, "author_name": "hugodeheer", "author_url": "", "post_date": "06/11/2024 12:38:23", "content": "Congratulations Sietse on your first silver medal!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2866265": "My solution is an ensemble of two high-scoring public notebooks:\n\nhttps://www.kaggle.com/code/tc0000/birdclef-starter-notebook (LB .654)\nhttps://www.kaggle.com/code/aikhmelnytskyy/birdclef24-pretraining-is-all-you-need-infer (LB .632)\n\nAll credit to their approaches, amazing work! \n\nSince 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. \n\nPersonally, 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 on private was minimal, showing the resilience of their approach to overfitting to public. Again, credits to the author(s) of this work!", "2866302": "Really good job! Upvoted!", "2866646": "Congratulations Sietse on your first silver medal!" }, "source": "meta" }