{
  "id": 117782,
  "title": "Validation dataset size",
  "url": "/competitions/pku-autonomous-driving/discussion/117782",
  "author_name": "",
  "post_date": "2019-11-17T20:29:48.073428800Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Just curious what people have found to be a good validation dataset size? The baseline CenterNet kernel is using just 1% of the data, I think that is a bit too small to have testing. But I dont want to make it too big because I want to give the model enough data to give a good fit estimation.</p>\n\n<p>What do you guys think? Maybe 10-20% is best?</p>",
  "messages": [
    {
      "id": "675271",
      "postDate": "11/17/2019 20:29:48",
      "content": "<p>Just curious what people have found to be a good validation dataset size? The baseline CenterNet kernel is using just 1% of the data, I think that is a bit too small to have testing. But I dont want to make it too big because I want to give the model enough data to give a good fit estimation.</p>\n\n<p>What do you guys think? Maybe 10-20% is best?</p>",
      "rawMarkdown": "Just curious what people have found to be a good validation dataset size? The baseline CenterNet kernel is using just 1% of the data, I think that is a bit too small to have testing. But I dont want to make it too big because I want to give the model enough data to give a good fit estimation.\n\nWhat do you guys think? Maybe 10-20% is best?",
      "votes": null
    },
    {
      "id": "675273",
      "postDate": "11/17/2019 20:35:41",
      "content": "<p>Validate on a proper sized holdout set and don't ever change it throughout your testing (or have a very good reason). I believe ~20 - 30% is OK in this case.</p>\n\n<p>If you have found a good solution, train with the exact same settings but on full train set and submit. </p>",
      "rawMarkdown": "Validate on a proper sized holdout set and don't ever change it throughout your testing (or have a very good reason). I believe ~20 - 30% is OK in this case.\n\nIf you have found a good solution, train with the exact same settings but on full train set and submit.",
      "votes": null
    },
    {
      "id": "675284",
      "postDate": "11/17/2019 21:00:43",
      "content": "<p>Agreed, I think I will increase the validation holdout and continue my testing from there.</p>",
      "rawMarkdown": "Agreed, I think I will increase the validation holdout and continue my testing from there.",
      "votes": null
    },
    {
      "id": "710879",
      "postDate": "01/05/2020 11:23:19",
      "content": "<p>so how many images you take as valiation?thanks~</p>",
      "rawMarkdown": "so how many images you take as valiation?thanks~",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 675273,
      "author_name": "ilu000",
      "author_url": "",
      "post_date": "11/17/2019 20:35:41",
      "content": "<p>Validate on a proper sized holdout set and don't ever change it throughout your testing (or have a very good reason). I believe ~20 - 30% is OK in this case.</p>\n\n<p>If you have found a good solution, train with the exact same settings but on full train set and submit. </p>",
      "votes": null,
      "replies": [
        {
          "id": 675284,
          "author_name": "danshatzz",
          "author_url": "",
          "post_date": "11/17/2019 21:00:43",
          "content": "<p>Agreed, I think I will increase the validation holdout and continue my testing from there.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 710879,
          "author_name": "hekaikai",
          "author_url": "",
          "post_date": "01/05/2020 11:23:19",
          "content": "<p>so how many images you take as valiation?thanks~</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "675271": "Just curious what people have found to be a good validation dataset size? The baseline CenterNet kernel is using just 1% of the data, I think that is a bit too small to have testing. But I dont want to make it too big because I want to give the model enough data to give a good fit estimation.\n\nWhat do you guys think? Maybe 10-20% is best?",
    "675273": "Validate on a proper sized holdout set and don't ever change it throughout your testing (or have a very good reason). I believe ~20 - 30% is OK in this case.\n\nIf you have found a good solution, train with the exact same settings but on full train set and submit.",
    "675284": "Agreed, I think I will increase the validation holdout and continue my testing from there.",
    "710879": "so how many images you take as valiation?thanks~"
  },
  "source": "meta"
}