{
  "id": 183218,
  "title": "My solution for my first medal (silver)",
  "url": "/competitions/birdsong-recognition/discussion/183218",
  "author_name": "yuvaramsingh",
  "post_date": "2020-09-16T01:26:32.686000",
  "votes": 5,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Thanks to kaggle and competition host for bringing this competition and congrats to the winners and other competitors. this is my first time working with audio data and i sure learned a lot about DL due to the nature of this competition. The domain mismatch between the training and testing environment posted a good challenge to constantly question my understanding about Data and DL understanding .</p>\n<p><strong>My solution:</strong><br>\nthis is my best solution which i did not select for final submission <strong>0.626</strong><br>\nThis is an ensemble of two best models from different fold and varying level of augmentation</p>\n<p>Feature extractor : <strong>Resnet18</strong><br>\nNetwork architecture : Sound Event Detection(Not PANN)<br>\nMethod of training : Multi instance learning<br>\nKey layer : <strong>Adaptive Auto pooling</strong> Presented <a href=\"https://arxiv.org/pdf/1804.10070.pdf\" target=\"_blank\">here</a><br>\nAugmentation: extracted 20 sec sound data and mixed with other randomly chosen audio sample, Low pass filter(as low frequency waves travel longer distance), added background noise extracted from the provided autio sample.<br>\nBest score threshold : 0.30</p>\n<p><strong>Things i should have focused:</strong><br>\ni completely ignored the fact that there is additional data available and posted in the competition. this would have given me some more benefits. not aware of domain adaptation techniques and hoping to learn from others solution. Not changing from Resnet18 base model. for some reason i kept on working on the data and ignored to change the extractor when i had the time to experiment with it.</p>\n<p>code: <a href=\"https://github.com/yuvaramsingh94/Cornell-Birdcall-Identification-kaggle-solution\" target=\"_blank\">github</a></p>\n<p>the code needs some more cleaning and i will add a detailed Readme </p>\n<p>Thanks to <a href=\"https://www.kaggle.com/hidehisaarai1213\" target=\"_blank\">@hidehisaarai1213</a> <a href=\"https://www.kaggle.com/ttahara\" target=\"_blank\">@ttahara</a> and others for constantly sharing ideas and findings to help others. <br>\nOn to the next competitions :)</p>",
  "messages": [
    {
      "id": 1012227,
      "postDate": "2020-09-16T01:26:32.687Z",
      "content": "<p>Thanks to kaggle and competition host for bringing this competition and congrats to the winners and other competitors. this is my first time working with audio data and i sure learned a lot about DL due to the nature of this competition. The domain mismatch between the training and testing environment posted a good challenge to constantly question my understanding about Data and DL understanding .</p>\n<p><strong>My solution:</strong><br>\nthis is my best solution which i did not select for final submission <strong>0.626</strong><br>\nThis is an ensemble of two best models from different fold and varying level of augmentation</p>\n<p>Feature extractor : <strong>Resnet18</strong><br>\nNetwork architecture : Sound Event Detection(Not PANN)<br>\nMethod of training : Multi instance learning<br>\nKey layer : <strong>Adaptive Auto pooling</strong> Presented <a href=\"https://arxiv.org/pdf/1804.10070.pdf\" target=\"_blank\">here</a><br>\nAugmentation: extracted 20 sec sound data and mixed with other randomly chosen audio sample, Low pass filter(as low frequency waves travel longer distance), added background noise extracted from the provided autio sample.<br>\nBest score threshold : 0.30</p>\n<p><strong>Things i should have focused:</strong><br>\ni completely ignored the fact that there is additional data available and posted in the competition. this would have given me some more benefits. not aware of domain adaptation techniques and hoping to learn from others solution. Not changing from Resnet18 base model. for some reason i kept on working on the data and ignored to change the extractor when i had the time to experiment with it.</p>\n<p>code: <a href=\"https://github.com/yuvaramsingh94/Cornell-Birdcall-Identification-kaggle-solution\" target=\"_blank\">github</a></p>\n<p>the code needs some more cleaning and i will add a detailed Readme </p>\n<p>Thanks to <a href=\"https://www.kaggle.com/hidehisaarai1213\" target=\"_blank\">@hidehisaarai1213</a> <a href=\"https://www.kaggle.com/ttahara\" target=\"_blank\">@ttahara</a> and others for constantly sharing ideas and findings to help others. <br>\nOn to the next competitions :)</p>",
      "rawMarkdown": "Thanks to kaggle and competition host for bringing this competition and congrats to the winners and other competitors. this is my first time working with audio data and i sure learned a lot about DL due to the nature of this competition. The domain mismatch between the training and testing environment posted a good challenge to constantly question my understanding about Data and DL understanding .\n\n**My solution:**\nthis is my best solution which i did not select for final submission **0.626**\nThis is an ensemble of two best models from different fold and varying level of augmentation\n\nFeature extractor : **Resnet18**\nNetwork architecture : Sound Event Detection(Not PANN)\nMethod of training : Multi instance learning\nKey layer : **Adaptive Auto pooling** Presented [here](https://arxiv.org/pdf/1804.10070.pdf)\nAugmentation: extracted 20 sec sound data and mixed with other randomly chosen audio sample, Low pass filter(as low frequency waves travel longer distance), added background noise extracted from the provided autio sample.\nBest score threshold : 0.30\n \n**Things i should have focused:**\ni completely ignored the fact that there is additional data available and posted in the competition. this would have given me some more benefits. not aware of domain adaptation techniques and hoping to learn from others solution. Not changing from Resnet18 base model. for some reason i kept on working on the data and ignored to change the extractor when i had the time to experiment with it.\n\ncode: [github](https://github.com/yuvaramsingh94/Cornell-Birdcall-Identification-kaggle-solution)\n\nthe code needs some more cleaning and i will add a detailed Readme \n\nThanks to @hidehisaarai1213 @ttahara and others for constantly sharing ideas and findings to help others. \nOn to the next competitions :)",
      "votes": 5
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1012227": "Thanks to kaggle and competition host for bringing this competition and congrats to the winners and other competitors. this is my first time working with audio data and i sure learned a lot about DL due to the nature of this competition. The domain mismatch between the training and testing environment posted a good challenge to constantly question my understanding about Data and DL understanding .\n\n**My solution:**\nthis is my best solution which i did not select for final submission **0.626**\nThis is an ensemble of two best models from different fold and varying level of augmentation\n\nFeature extractor : **Resnet18**\nNetwork architecture : Sound Event Detection(Not PANN)\nMethod of training : Multi instance learning\nKey layer : **Adaptive Auto pooling** Presented [here](https://arxiv.org/pdf/1804.10070.pdf)\nAugmentation: extracted 20 sec sound data and mixed with other randomly chosen audio sample, Low pass filter(as low frequency waves travel longer distance), added background noise extracted from the provided autio sample.\nBest score threshold : 0.30\n \n**Things i should have focused:**\ni completely ignored the fact that there is additional data available and posted in the competition. this would have given me some more benefits. not aware of domain adaptation techniques and hoping to learn from others solution. Not changing from Resnet18 base model. for some reason i kept on working on the data and ignored to change the extractor when i had the time to experiment with it.\n\ncode: [github](https://github.com/yuvaramsingh94/Cornell-Birdcall-Identification-kaggle-solution)\n\nthe code needs some more cleaning and i will add a detailed Readme \n\nThanks to @hidehisaarai1213 @ttahara and others for constantly sharing ideas and findings to help others. \nOn to the next competitions :)"
  }
}