{
  "id": 308004,
  "title": "First submission(BirdCLEF 2021 4th solution)",
  "url": "/competitions/birdclef-2022/discussion/308004",
  "author_name": "",
  "post_date": "2022-02-16T16:53:44.456895200Z",
  "votes": 51,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I don't know if I will be able to spend much time on BirdCLEF2022, but I share my results using BirdCLEF2021 4th solution.<br>\ntrain code: <a href=\"https://github.com/tattaka/birdclef-2021\" target=\"_blank\">https://github.com/tattaka/birdclef-2021</a><br>\nnotebook(not clean): <a href=\"https://www.kaggle.com/tattaka/birdclef2022-submission-baseline\" target=\"_blank\">https://www.kaggle.com/tattaka/birdclef2022-submission-baseline</a></p>\n<h3>impressions</h3>\n<ul>\n<li>I used all train_audio for learning.<ul>\n<li>I am going to extract only scored_birds in the next step.</li></ul></li>\n<li>Inference takes a lot of time because of the large amount of test data.<ul>\n<li>I took 2 hours to infer the above notebook (there are much unnecessary processing, so there is room for improvement).</li>\n<li>Since ensembles are often effective in audio competitions, I expect more other unique solutions 😉.</li></ul></li>\n</ul>",
  "messages": [
    {
      "id": "1693410",
      "postDate": "02/16/2022 16:53:44",
      "content": "<p>I don't know if I will be able to spend much time on BirdCLEF2022, but I share my results using BirdCLEF2021 4th solution.<br>\ntrain code: <a href=\"https://github.com/tattaka/birdclef-2021\" target=\"_blank\">https://github.com/tattaka/birdclef-2021</a><br>\nnotebook(not clean): <a href=\"https://www.kaggle.com/tattaka/birdclef2022-submission-baseline\" target=\"_blank\">https://www.kaggle.com/tattaka/birdclef2022-submission-baseline</a></p>\n<h3>impressions</h3>\n<ul>\n<li>I used all train_audio for learning.<ul>\n<li>I am going to extract only scored_birds in the next step.</li></ul></li>\n<li>Inference takes a lot of time because of the large amount of test data.<ul>\n<li>I took 2 hours to infer the above notebook (there are much unnecessary processing, so there is room for improvement).</li>\n<li>Since ensembles are often effective in audio competitions, I expect more other unique solutions 😉.</li></ul></li>\n</ul>",
      "rawMarkdown": "I don't know if I will be able to spend much time on BirdCLEF2022, but I share my results using BirdCLEF2021 4th solution.\ntrain code: https://github.com/tattaka/birdclef-2021\nnotebook(not clean): https://www.kaggle.com/tattaka/birdclef2022-submission-baseline\n\n### impressions\n* I used all train_audio for learning.\n    *  I am going to extract only scored_birds in the next step.\n* Inference takes a lot of time because of the large amount of test data.\n    * I took 2 hours to infer the above notebook (there are much unnecessary processing, so there is room for improvement).\n    * Since ensembles are often effective in audio competitions, I expect more other unique solutions 😉.",
      "votes": null
    },
    {
      "id": "1693474",
      "postDate": "02/16/2022 17:37:02",
      "content": "<p>Thanks for sharing and nice to see the previous competition solutions can be part of this competition too</p>\n<p><a href=\"https://www.kaggle.com/tattaka\" target=\"_blank\">@tattaka</a> how long to train all your models for stage1, stage2 and final </p>",
      "rawMarkdown": "Thanks for sharing and nice to see the previous competition solutions can be part of this competition too\n\n@tattaka how long to train all your models for stage1, stage2 and final",
      "votes": null
    },
    {
      "id": "1693704",
      "postDate": "02/16/2022 22:21:56",
      "content": "<p>Stage1 only, 50epoch</p>",
      "rawMarkdown": "Stage1 only, 50epoch",
      "votes": null
    },
    {
      "id": "1694002",
      "postDate": "02/17/2022 05:35:01",
      "content": "<p>Great work <a href=\"https://www.kaggle.com/tattaka\" target=\"_blank\">@tattaka</a> 😎Thanks for sharing!🤗 new follower 🙋‍♀️</p>",
      "rawMarkdown": "Great work @tattaka 😎Thanks for sharing!🤗 new follower 🙋‍♀️",
      "votes": null
    },
    {
      "id": "1696373",
      "postDate": "02/18/2022 19:34:55",
      "content": "<p>Thanks for sharing ! I'm trying to reproduce the training with your training code from the github repo. Did you train with precision=16 with lightning Trainer since you are using amp.autocast in inference ? Unfortunately, I got Nan loss when using precision=16 and resnest26d with your training code (BS=36, lr=1e-3, backbone_lr=1e-4, period=30.0, infer_period=30.0)</p>",
      "rawMarkdown": "Thanks for sharing ! I'm trying to reproduce the training with your training code from the github repo. Did you train with precision=16 with lightning Trainer since you are using amp.autocast in inference ? Unfortunately, I got Nan loss when using precision=16 and resnest26d with your training code (BS=36, lr=1e-3, backbone_lr=1e-4, period=30.0, infer_period=30.0)",
      "votes": null
    },
    {
      "id": "1696539",
      "postDate": "02/18/2022 23:11:31",
      "content": "<p>I didn't train with precision=16. The inference code amp.autocast is unnecessary.I forgot to remove it.</p>",
      "rawMarkdown": "I didn't train with precision=16. The inference code amp.autocast is unnecessary.I forgot to remove it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1693474,
      "author_name": "seshurajup",
      "author_url": "",
      "post_date": "02/16/2022 17:37:02",
      "content": "<p>Thanks for sharing and nice to see the previous competition solutions can be part of this competition too</p>\n<p><a href=\"https://www.kaggle.com/tattaka\" target=\"_blank\">@tattaka</a> how long to train all your models for stage1, stage2 and final </p>",
      "votes": null,
      "replies": [
        {
          "id": 1693704,
          "author_name": "tattaka",
          "author_url": "",
          "post_date": "02/16/2022 22:21:56",
          "content": "<p>Stage1 only, 50epoch</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1694002,
      "author_name": "arunasivapragasam",
      "author_url": "",
      "post_date": "02/17/2022 05:35:01",
      "content": "<p>Great work <a href=\"https://www.kaggle.com/tattaka\" target=\"_blank\">@tattaka</a> 😎Thanks for sharing!🤗 new follower 🙋‍♀️</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1696373,
      "author_name": "alexandrecc",
      "author_url": "",
      "post_date": "02/18/2022 19:34:55",
      "content": "<p>Thanks for sharing ! I'm trying to reproduce the training with your training code from the github repo. Did you train with precision=16 with lightning Trainer since you are using amp.autocast in inference ? Unfortunately, I got Nan loss when using precision=16 and resnest26d with your training code (BS=36, lr=1e-3, backbone_lr=1e-4, period=30.0, infer_period=30.0)</p>",
      "votes": null,
      "replies": [
        {
          "id": 1696539,
          "author_name": "tattaka",
          "author_url": "",
          "post_date": "02/18/2022 23:11:31",
          "content": "<p>I didn't train with precision=16. The inference code amp.autocast is unnecessary.I forgot to remove it.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1693410": "I don't know if I will be able to spend much time on BirdCLEF2022, but I share my results using BirdCLEF2021 4th solution.\ntrain code: https://github.com/tattaka/birdclef-2021\nnotebook(not clean): https://www.kaggle.com/tattaka/birdclef2022-submission-baseline\n\n### impressions\n* I used all train_audio for learning.\n    *  I am going to extract only scored_birds in the next step.\n* Inference takes a lot of time because of the large amount of test data.\n    * I took 2 hours to infer the above notebook (there are much unnecessary processing, so there is room for improvement).\n    * Since ensembles are often effective in audio competitions, I expect more other unique solutions 😉.",
    "1693474": "Thanks for sharing and nice to see the previous competition solutions can be part of this competition too\n\n@tattaka how long to train all your models for stage1, stage2 and final",
    "1693704": "Stage1 only, 50epoch",
    "1694002": "Great work @tattaka 😎Thanks for sharing!🤗 new follower 🙋‍♀️",
    "1696373": "Thanks for sharing ! I'm trying to reproduce the training with your training code from the github repo. Did you train with precision=16 with lightning Trainer since you are using amp.autocast in inference ? Unfortunately, I got Nan loss when using precision=16 and resnest26d with your training code (BS=36, lr=1e-3, backbone_lr=1e-4, period=30.0, infer_period=30.0)",
    "1696539": "I didn't train with precision=16. The inference code amp.autocast is unnecessary.I forgot to remove it."
  },
  "source": "meta"
}