{
  "id": 69341,
  "title": "Thresholding predictions from the model",
  "url": "/competitions/inclusive-images-challenge/discussion/69341",
  "author_name": "",
  "post_date": "2018-10-23T01:26:08.473982500Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi,</p>\n\n<p>What would be the best way to choose which classes to assign to an image after getting the outputs from the network? Would thresholding predictions at 0.1 or 0.01 be best, or should you choose the top 3 or top k classes?</p>\n\n<p>Thanks,</p>\n\n<p>Bilal</p>",
  "messages": [
    {
      "id": "408486",
      "postDate": "10/23/2018 01:26:08",
      "content": "<p>Hi,</p>\n\n<p>What would be the best way to choose which classes to assign to an image after getting the outputs from the network? Would thresholding predictions at 0.1 or 0.01 be best, or should you choose the top 3 or top k classes?</p>\n\n<p>Thanks,</p>\n\n<p>Bilal</p>",
      "rawMarkdown": "Hi,\n\nWhat would be the best way to choose which classes to assign to an image after getting the outputs from the network? Would thresholding predictions at 0.1 or 0.01 be best, or should you choose the top 3 or top k classes?\n\nThanks,\n\nBilal",
      "votes": null
    },
    {
      "id": "410556",
      "postDate": "10/26/2018 08:53:11",
      "content": "<p>Usually, we choose threshold by validation. You pick up to 10 thresholds in a range and then see which one gets best F2 score.</p>",
      "rawMarkdown": "Usually, we choose threshold by validation. You pick up to 10 thresholds in a range and then see which one gets best F2 score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 410556,
      "author_name": "whilefalse",
      "author_url": "",
      "post_date": "10/26/2018 08:53:11",
      "content": "<p>Usually, we choose threshold by validation. You pick up to 10 thresholds in a range and then see which one gets best F2 score.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "408486": "Hi,\n\nWhat would be the best way to choose which classes to assign to an image after getting the outputs from the network? Would thresholding predictions at 0.1 or 0.01 be best, or should you choose the top 3 or top k classes?\n\nThanks,\n\nBilal",
    "410556": "Usually, we choose threshold by validation. You pick up to 10 thresholds in a range and then see which one gets best F2 score."
  },
  "source": "meta"
}