I realized that we must check 3 aspects of this competition to score up
\nIt goes without saying that we have to choose good models and get good weights to improve the cross-validation score, and I think TimmSED(PANNs model replaced by Timm backbone network) introduced in these notebooks is one of the good models.
\n[PyTorch, Training] BirdCLEF2021 Starter
\n[BirdCLEF2022] EX005 f0 [Infer]
Second, we should watch out for domain shift, which is the difference in the distribution between train data and test data. As you know, the sound collection environment of train data is different from that of test data(test data is supposed to be noisier). I think the key to eliminating domain shift is noise injection(I also commented about noise injection in this discussion).
\nFinally, we should choose the appropriate threshold to score up. @shinmurashinmura introduces great idea about this topic. Check it out↓
\nThe lower threshold has good score.
\nVisualization soundscape
In my experience, the appropriate threshold can be changed drastically by minor parameter tuning.
", "messages": [ { "id": "1770181", "postDate": "04/28/2022 04:19:26", "content": "I realized that we must check 3 aspects of this competition to score up
\nIt goes without saying that we have to choose good models and get good weights to improve the cross-validation score, and I think TimmSED(PANNs model replaced by Timm backbone network) introduced in these notebooks is one of the good models.
\n[PyTorch, Training] BirdCLEF2021 Starter
\n[BirdCLEF2022] EX005 f0 [Infer]
Second, we should watch out for domain shift, which is the difference in the distribution between train data and test data. As you know, the sound collection environment of train data is different from that of test data(test data is supposed to be noisier). I think the key to eliminating domain shift is noise injection(I also commented about noise injection in this discussion).
\nFinally, we should choose the appropriate threshold to score up. @shinmurashinmura introduces great idea about this topic. Check it out↓
\nThe lower threshold has good score.
\nVisualization soundscape
In my experience, the appropriate threshold can be changed drastically by minor parameter tuning.
", "rawMarkdown": "I realized that we must check 3 aspects of this competition to score up\n\n* Improve cross-validation score.\n* Eliminate the effect of domain shift.\n* Choose an appropriate threshold.\n\nIt goes without saying that we have to choose good models and get good weights to improve the cross-validation score, and I think TimmSED(PANNs model replaced by Timm backbone network) introduced in these notebooks is one of the good models.\n\n[[PyTorch, Training] BirdCLEF2021 Starter](https://www.kaggle.com/code/hidehisaarai1213/pytorch-training-birdclef2021-starter)\n[[BirdCLEF2022] EX005 f0 [Infer]](https://www.kaggle.com/code/kaerunantoka/birdclef2022-ex005-f0-infer)\n\nSecond, we should watch out for domain shift, which is the difference in the distribution between train data and test data. As you know, the sound collection environment of train data is different from that of test data(test data is supposed to be noisier). I think the key to eliminating domain shift is noise injection(I also commented about noise injection in [this discussion](https://www.kaggle.com/competitions/birdclef-2022/discussion/320613)).\n\nFinally, we should choose the appropriate threshold to score up. @shinmurashinmura introduces great idea about this topic. Check it out↓\n[The lower threshold has good score.](https://www.kaggle.com/competitions/birdclef-2022/discussion/318999)\n[Visualization soundscape](https://www.kaggle.com/competitions/birdclef-2022/discussion/320304)\n\nIn my experience, the appropriate threshold can be changed drastically by minor parameter tuning.", "votes": null }, { "id": "1780894", "postDate": "05/08/2022 02:11:40", "content": "Nice work.
", "rawMarkdown": "Nice work.", "votes": null }, { "id": "1792363", "postDate": "05/16/2022 22:31:08", "content": "This is very helpful!
", "rawMarkdown": "This is very helpful!", "votes": null } ], "comments": [ { "id": 1780894, "author_name": "lemelo", "author_url": "", "post_date": "05/08/2022 02:11:40", "content": "Nice work.
", "votes": null, "replies": [] }, { "id": 1792363, "author_name": "bhavesh0124", "author_url": "", "post_date": "05/16/2022 22:31:08", "content": "This is very helpful!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1770181": "I realized that we must check 3 aspects of this competition to score up\n\n* Improve cross-validation score.\n* Eliminate the effect of domain shift.\n* Choose an appropriate threshold.\n\nIt goes without saying that we have to choose good models and get good weights to improve the cross-validation score, and I think TimmSED(PANNs model replaced by Timm backbone network) introduced in these notebooks is one of the good models.\n\n[[PyTorch, Training] BirdCLEF2021 Starter](https://www.kaggle.com/code/hidehisaarai1213/pytorch-training-birdclef2021-starter)\n[[BirdCLEF2022] EX005 f0 [Infer]](https://www.kaggle.com/code/kaerunantoka/birdclef2022-ex005-f0-infer)\n\nSecond, we should watch out for domain shift, which is the difference in the distribution between train data and test data. As you know, the sound collection environment of train data is different from that of test data(test data is supposed to be noisier). I think the key to eliminating domain shift is noise injection(I also commented about noise injection in [this discussion](https://www.kaggle.com/competitions/birdclef-2022/discussion/320613)).\n\nFinally, we should choose the appropriate threshold to score up. @shinmurashinmura introduces great idea about this topic. Check it out↓\n[The lower threshold has good score.](https://www.kaggle.com/competitions/birdclef-2022/discussion/318999)\n[Visualization soundscape](https://www.kaggle.com/competitions/birdclef-2022/discussion/320304)\n\nIn my experience, the appropriate threshold can be changed drastically by minor parameter tuning.", "1780894": "Nice work.", "1792363": "This is very helpful!" }, "source": "meta" }