{
  "id": 182502,
  "title": "Inference Hacks",
  "url": "/competitions/birdsong-recognition/discussion/182502",
  "author_name": "DhruvBishnoi",
  "post_date": "2020-09-13T05:54:13.869000",
  "votes": 0,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi Everyone, <br>\nTried to look for <strong>THE MAGIC</strong> in the discussions forum but didn't find any. </p>\n<p>I am currently using PANNs. Here are some ideas : </p>\n<ol>\n<li><p>SITE_1, SITE_2 are to be predicted every five seconds interval. Some of the bird calls are overlapping between 2 five seconds window. For example, a bird call starts at 19 sec and ends at 21 sec.  In which 5-sec bucket should I label the bird? <em>a. 15-20 sec only  b. 20-25 sec only  c. 15-20 and 20-25 sec both</em></p></li>\n<li><p>Rounding start time and end time as per 5 sec, For example, start time: 19.84 -&gt; 20 or end time: 25.10 -&gt; 25.  </p></li>\n<li><p>When the model predicts multiple overlapping predictions, it could be a no-call as well.</p></li>\n</ol>\n<p>Hoping people to share some hints on post-processing. </p>",
  "messages": [
    {
      "id": 1008550,
      "postDate": "2020-09-13T07:52:36.943Z",
      "content": "<p>I don't do any post-processing at all. <br>\nMaybe some top competitors use some 'magic' but so far I can see only that most teams just don't do well <strong>domain shift</strong>. And that's very real production problem by the way. </p>",
      "rawMarkdown": "I don't do any post-processing at all. \nMaybe some top competitors use some 'magic' but so far I can see only that most teams just don't do well **domain shift**. And that's very real production problem by the way. ",
      "votes": 10
    },
    {
      "id": 1008556,
      "postDate": "2020-09-13T08:05:44.450Z",
      "content": "<blockquote>\n  <p>Tried to look for THE MAGIC in the discussions forum but didn't find any.</p>\n</blockquote>\n<p>There is no magic that we are aware of, or leak or whatever you want to call it. </p>\n<p>And even if there was magic that would be a pretty bad idea for anyone to reveal it 2 days before the end of the competition.</p>",
      "rawMarkdown": "> Tried to look for THE MAGIC in the discussions forum but didn't find any.\n\nThere is no magic that we are aware of, or leak or whatever you want to call it. \n\nAnd even if there was magic that would be a pretty bad idea for anyone to reveal it 2 days before the end of the competition.",
      "votes": 5
    },
    {
      "id": 1008437,
      "postDate": "2020-09-13T05:54:13.870Z",
      "content": "<p>Hi Everyone, <br>\nTried to look for <strong>THE MAGIC</strong> in the discussions forum but didn't find any. </p>\n<p>I am currently using PANNs. Here are some ideas : </p>\n<ol>\n<li><p>SITE_1, SITE_2 are to be predicted every five seconds interval. Some of the bird calls are overlapping between 2 five seconds window. For example, a bird call starts at 19 sec and ends at 21 sec.  In which 5-sec bucket should I label the bird? <em>a. 15-20 sec only  b. 20-25 sec only  c. 15-20 and 20-25 sec both</em></p></li>\n<li><p>Rounding start time and end time as per 5 sec, For example, start time: 19.84 -&gt; 20 or end time: 25.10 -&gt; 25.  </p></li>\n<li><p>When the model predicts multiple overlapping predictions, it could be a no-call as well.</p></li>\n</ol>\n<p>Hoping people to share some hints on post-processing. </p>",
      "rawMarkdown": "Hi Everyone, \nTried to look for **THE MAGIC** in the discussions forum but didn't find any. \n\nI am currently using PANNs. Here are some ideas : \n1. SITE_1, SITE_2 are to be predicted every five seconds interval. Some of the bird calls are overlapping between 2 five seconds window. For example, a bird call starts at 19 sec and ends at 21 sec.  In which 5-sec bucket should I label the bird? *a. 15-20 sec only  b. 20-25 sec only  c. 15-20 and 20-25 sec both*\n\n2. Rounding start time and end time as per 5 sec, For example, start time: 19.84 -> 20 or end time: 25.10 -> 25.  \n3. When the model predicts multiple overlapping predictions, it could be a no-call as well.\n \n\nHoping people to share some hints on post-processing. \n"
    },
    {
      "id": 1009117,
      "postDate": "2020-09-13T16:57:34.933Z",
      "rawMarkdown": "",
      "votes": -1,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1008550,
      "author_name": "Oleg Yaroshevskiy",
      "author_url": "",
      "post_date": "2020-09-13T07:52:36.943000",
      "content": "<p>I don't do any post-processing at all. <br>\nMaybe some top competitors use some 'magic' but so far I can see only that most teams just don't do well <strong>domain shift</strong>. And that's very real production problem by the way. </p>",
      "votes": 10,
      "replies": []
    },
    {
      "id": 1008556,
      "author_name": "Theo Viel",
      "author_url": "",
      "post_date": "2020-09-13T08:05:44.450000",
      "content": "<blockquote>\n  <p>Tried to look for THE MAGIC in the discussions forum but didn't find any.</p>\n</blockquote>\n<p>There is no magic that we are aware of, or leak or whatever you want to call it. </p>\n<p>And even if there was magic that would be a pretty bad idea for anyone to reveal it 2 days before the end of the competition.</p>",
      "votes": 5,
      "replies": []
    },
    {
      "id": 1009117,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-09-13T16:57:34.933000",
      "content": "",
      "votes": -1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1008550": "I don't do any post-processing at all. \nMaybe some top competitors use some 'magic' but so far I can see only that most teams just don't do well **domain shift**. And that's very real production problem by the way. ",
    "1008556": "> Tried to look for THE MAGIC in the discussions forum but didn't find any.\n\nThere is no magic that we are aware of, or leak or whatever you want to call it. \n\nAnd even if there was magic that would be a pretty bad idea for anyone to reveal it 2 days before the end of the competition.",
    "1008437": "Hi Everyone, \nTried to look for **THE MAGIC** in the discussions forum but didn't find any. \n\nI am currently using PANNs. Here are some ideas : \n1. SITE_1, SITE_2 are to be predicted every five seconds interval. Some of the bird calls are overlapping between 2 five seconds window. For example, a bird call starts at 19 sec and ends at 21 sec.  In which 5-sec bucket should I label the bird? *a. 15-20 sec only  b. 20-25 sec only  c. 15-20 and 20-25 sec both*\n\n2. Rounding start time and end time as per 5 sec, For example, start time: 19.84 -> 20 or end time: 25.10 -> 25.  \n3. When the model predicts multiple overlapping predictions, it could be a no-call as well.\n \n\nHoping people to share some hints on post-processing. \n",
    "1009117": ""
  }
}