{
  "id": 183610,
  "title": "Just an idea",
  "url": "/competitions/landmark-recognition-2020/discussion/183610",
  "author_name": "",
  "post_date": "2020-09-17T12:25:08.895612Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I am not much experienced and not well versed with the libraries as well but I have an idea for a model I wanna share. Any suggestion/ criticism is welcome.<br>\nI am thinking something along the lines of having an encoder and encoding the images into a feature vector. But since the number of classes is huge ~81k and samples for a few of them are sparse so a decoder for each class sounds impractical. What I thought instead is to encode each class into a strict number sequence(1,7,9,…) -&gt; (1,2,3,…) and then converting these to binary. Binary will be length 17 so we train 17 binary classifier and this way we should have ample data for each class.<br>\nAgain, This is just my idea. I am a total noob so could be very wrong. In that case, please point out the mistakes I made. Thanks in advance.</p>",
  "messages": [
    {
      "id": "1014412",
      "postDate": "09/17/2020 12:25:08",
      "content": "<p>I am not much experienced and not well versed with the libraries as well but I have an idea for a model I wanna share. Any suggestion/ criticism is welcome.<br>\nI am thinking something along the lines of having an encoder and encoding the images into a feature vector. But since the number of classes is huge ~81k and samples for a few of them are sparse so a decoder for each class sounds impractical. What I thought instead is to encode each class into a strict number sequence(1,7,9,…) -&gt; (1,2,3,…) and then converting these to binary. Binary will be length 17 so we train 17 binary classifier and this way we should have ample data for each class.<br>\nAgain, This is just my idea. I am a total noob so could be very wrong. In that case, please point out the mistakes I made. Thanks in advance.</p>",
      "rawMarkdown": "I am not much experienced and not well versed with the libraries as well but I have an idea for a model I wanna share. Any suggestion/ criticism is welcome.\nI am thinking something along the lines of having an encoder and encoding the images into a feature vector. But since the number of classes is huge ~81k and samples for a few of them are sparse so a decoder for each class sounds impractical. What I thought instead is to encode each class into a strict number sequence(1,7,9,...) -> (1,2,3,...) and then converting these to binary. Binary will be length 17 so we train 17 binary classifier and this way we should have ample data for each class.\nAgain, This is just my idea. I am a total noob so could be very wrong. In that case, please point out the mistakes I made. Thanks in advance.",
      "votes": null
    },
    {
      "id": "1014624",
      "postDate": "09/17/2020 15:46:59",
      "content": "<p>Okay exploding to binary might be bad but some sort of clustering depending on scenery for example mountains and lakes to be separated etc beforehand sounds really good</p>",
      "rawMarkdown": "Okay exploding to binary might be bad but some sort of clustering depending on scenery for example mountains and lakes to be separated etc beforehand sounds really good",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1014624,
      "author_name": "washerman",
      "author_url": "",
      "post_date": "09/17/2020 15:46:59",
      "content": "<p>Okay exploding to binary might be bad but some sort of clustering depending on scenery for example mountains and lakes to be separated etc beforehand sounds really good</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1014412": "I am not much experienced and not well versed with the libraries as well but I have an idea for a model I wanna share. Any suggestion/ criticism is welcome.\nI am thinking something along the lines of having an encoder and encoding the images into a feature vector. But since the number of classes is huge ~81k and samples for a few of them are sparse so a decoder for each class sounds impractical. What I thought instead is to encode each class into a strict number sequence(1,7,9,...) -> (1,2,3,...) and then converting these to binary. Binary will be length 17 so we train 17 binary classifier and this way we should have ample data for each class.\nAgain, This is just my idea. I am a total noob so could be very wrong. In that case, please point out the mistakes I made. Thanks in advance.",
    "1014624": "Okay exploding to binary might be bad but some sort of clustering depending on scenery for example mountains and lakes to be separated etc beforehand sounds really good"
  },
  "source": "meta"
}