{
  "id": 133203,
  "title": "How many folds do you use in cross validation?",
  "url": "/competitions/bengaliai-cv19/discussion/133203",
  "author_name": "",
  "post_date": "2020-03-01T09:59:38.405019700Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>How many folds do you use in cross validation? Thanks!</p>",
  "messages": [
    {
      "id": "760447",
      "postDate": "03/01/2020 09:59:38",
      "content": "<p>How many folds do you use in cross validation? Thanks!</p>",
      "rawMarkdown": "How many folds do you use in cross validation? Thanks!",
      "votes": null
    },
    {
      "id": "760451",
      "postDate": "03/01/2020 10:06:29",
      "content": "<p>I started at 3, but that doesn't give much data for training.</p>\n\n<p>Since, given N the number of fold, you train on <code>number_of_images / (N-1)</code> images, if you have N=3, you train on 67% of your data. So I'm working on 5 for my part now (80% of the data per fold).</p>\n\n<p>It requires 5x longer training than a single 80/20 fold, so I try doing experiments on 1 single fold (stratified on <code>grapheme_roots</code> for experiments. Then, after finding something useful, I apply it to the 5 Fold CV run.</p>\n\n<p>Hope that helps!</p>",
      "rawMarkdown": "I started at 3, but that doesn't give much data for training.\n\nSince, given N the number of fold, you train on `number_of_images / (N-1)` images, if you have N=3, you train on 67% of your data. So I'm working on 5 for my part now (80% of the data per fold).\n\nIt requires 5x longer training than a single 80/20 fold, so I try doing experiments on 1 single fold (stratified on `grapheme_roots` for experiments. Then, after finding something useful, I apply it to the 5 Fold CV run.\n\nHope that helps!",
      "votes": null
    },
    {
      "id": "761001",
      "postDate": "03/02/2020 02:40:44",
      "content": "<p>Thanks!</p>",
      "rawMarkdown": "Thanks!",
      "votes": null
    },
    {
      "id": "767755",
      "postDate": "03/10/2020 04:02:29",
      "content": "<p>Thanks for sharing</p>",
      "rawMarkdown": "Thanks for sharing",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 760451,
      "author_name": "maxlenormand",
      "author_url": "",
      "post_date": "03/01/2020 10:06:29",
      "content": "<p>I started at 3, but that doesn't give much data for training.</p>\n\n<p>Since, given N the number of fold, you train on <code>number_of_images / (N-1)</code> images, if you have N=3, you train on 67% of your data. So I'm working on 5 for my part now (80% of the data per fold).</p>\n\n<p>It requires 5x longer training than a single 80/20 fold, so I try doing experiments on 1 single fold (stratified on <code>grapheme_roots</code> for experiments. Then, after finding something useful, I apply it to the 5 Fold CV run.</p>\n\n<p>Hope that helps!</p>",
      "votes": null,
      "replies": [
        {
          "id": 761001,
          "author_name": "jupanlee12",
          "author_url": "",
          "post_date": "03/02/2020 02:40:44",
          "content": "<p>Thanks!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 767755,
          "author_name": "kurianbenoy",
          "author_url": "",
          "post_date": "03/10/2020 04:02:29",
          "content": "<p>Thanks for sharing</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "760447": "How many folds do you use in cross validation? Thanks!",
    "760451": "I started at 3, but that doesn't give much data for training.\n\nSince, given N the number of fold, you train on `number_of_images / (N-1)` images, if you have N=3, you train on 67% of your data. So I'm working on 5 for my part now (80% of the data per fold).\n\nIt requires 5x longer training than a single 80/20 fold, so I try doing experiments on 1 single fold (stratified on `grapheme_roots` for experiments. Then, after finding something useful, I apply it to the 5 Fold CV run.\n\nHope that helps!",
    "761001": "Thanks!",
    "767755": "Thanks for sharing"
  },
  "source": "meta"
}