{
  "id": 71823,
  "title": "How do you use the information \"recognized\"?",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/71823",
  "author_name": "Youhan Lee",
  "post_date": "2018-11-17T03:03:02.208000",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>As you know, the csv files contain the information of \"recongnized\". </p>\n\n<p>I'm using the information to my model as deepsupervision or additional channel.</p>\n\n<p>How do you, smart kagglers, use it?</p>",
  "messages": [
    {
      "id": 427311,
      "postDate": "2018-11-25T06:56:50.317Z",
      "content": "<p>I set the wight of recognized=True samples to 1.0 and recognized=False samples to 0.5, because it make sense to set crapy samples a lower weight.</p>",
      "rawMarkdown": "I set the wight of recognized=True samples to 1.0 and recognized=False samples to 0.5, because it make sense to set crapy samples a lower weight.",
      "votes": 1,
      "replies": [
        {
          "id": 430969,
          "postDate": "2018-12-01T10:53:31.360Z",
          "content": "<p>Did that help with the score? </p>",
          "rawMarkdown": "Did that help with the score? ",
          "votes": 2
        },
        {
          "id": 431428,
          "postDate": "2018-12-02T08:50:37.260Z",
          "content": "<p>Actually I don't see significant improvement. There are 4194827 <code>recognized=False</code> samples, which is only 8.4% in the whole dataset, the ratio of <code>recognized=False</code> samples is too small, I think this is why sample weight is negligible</p>",
          "rawMarkdown": "Actually I don't see significant improvement. There are 4194827 `recognized=False` samples, which is only 8.4% in the whole dataset, the ratio of `recognized=False` samples is too small, I think this is why sample weight is negligible"
        }
      ]
    },
    {
      "id": 422909,
      "postDate": "2018-11-17T03:03:02.210Z",
      "content": "<p>As you know, the csv files contain the information of \"recongnized\". </p>\n\n<p>I'm using the information to my model as deepsupervision or additional channel.</p>\n\n<p>How do you, smart kagglers, use it?</p>",
      "rawMarkdown": "As you know, the csv files contain the information of \"recongnized\". \n\nI'm using the information to my model as deepsupervision or additional channel.\n\nHow do you, smart kagglers, use it?",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 427311,
      "author_name": "[he.ai]soulmachine",
      "author_url": "",
      "post_date": "2018-11-25T06:56:50.317000",
      "content": "<p>I set the wight of recognized=True samples to 1.0 and recognized=False samples to 0.5, because it make sense to set crapy samples a lower weight.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 430969,
          "author_name": "HuyenNguyen",
          "author_url": "",
          "post_date": "2018-12-01T10:53:31.360000",
          "content": "<p>Did that help with the score? </p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 431428,
          "author_name": "[he.ai]soulmachine",
          "author_url": "",
          "post_date": "2018-12-02T08:50:37.260000",
          "content": "<p>Actually I don't see significant improvement. There are 4194827 <code>recognized=False</code> samples, which is only 8.4% in the whole dataset, the ratio of <code>recognized=False</code> samples is too small, I think this is why sample weight is negligible</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "427311": "I set the wight of recognized=True samples to 1.0 and recognized=False samples to 0.5, because it make sense to set crapy samples a lower weight.",
    "422909": "As you know, the csv files contain the information of \"recongnized\". \n\nI'm using the information to my model as deepsupervision or additional channel.\n\nHow do you, smart kagglers, use it?"
  }
}