{
  "id": 112780,
  "title": "CV & LB corelation.",
  "url": "/competitions/understanding_cloud_organization/discussion/112780",
  "author_name": "",
  "post_date": "2019-10-15T10:26:09.561874900Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello. I see some kernels, and in many of them val_dice_coef is ~0.55, but LB score is 0.645.</p>\n\n<p>How i can see my local score, which will be corelation with LB score.</p>",
  "messages": [
    {
      "id": "649420",
      "postDate": "10/15/2019 10:26:09",
      "content": "<p>Hello. I see some kernels, and in many of them val_dice_coef is ~0.55, but LB score is 0.645.</p>\n\n<p>How i can see my local score, which will be corelation with LB score.</p>",
      "rawMarkdown": "Hello. I see some kernels, and in many of them val_dice_coef is ~0.55, but LB score is 0.645.\n\nHow i can see my local score, which will be corelation with LB score.",
      "votes": null
    },
    {
      "id": "649423",
      "postDate": "10/15/2019 10:33:36",
      "content": "<p>I fix this by not applying smooth to my dice coef metric.</p>",
      "rawMarkdown": "I fix this by not applying smooth to my dice coef metric.",
      "votes": null
    },
    {
      "id": "659342",
      "postDate": "10/27/2019 13:42:33",
      "content": "<p>Kaggle computes dice for each pair of true and predicted and then averages them all. If you use the popular validation dice coef code in public notebooks, it doesn't do that. The correct metric to compile into your model is posted <a href=\"https://www.kaggle.com/c/understanding_cloud_organization/discussion/114093\">here</a>. Use the metric named <code>mean_dice_coef</code> in your models. (It uses the other function <code>single_dice_coef</code> but you don't use that one).</p>\n\n<p>There are other reasons why CV and LB may not correlate, but the above makes your model use the same scoring mechanism that Kaggle does.</p>",
      "rawMarkdown": "Kaggle computes dice for each pair of true and predicted and then averages them all. If you use the popular validation dice coef code in public notebooks, it doesn't do that. The correct metric to compile into your model is posted [here][1]. Use the metric named `mean_dice_coef` in your models. (It uses the other function `single_dice_coef` but you don't use that one).\n\nThere are other reasons why CV and LB may not correlate, but the above makes your model use the same scoring mechanism that Kaggle does.\n\n[1]: https://www.kaggle.com/c/understanding_cloud_organization/discussion/114093",
      "votes": null
    },
    {
      "id": "659413",
      "postDate": "10/27/2019 16:10:00",
      "content": "<p>I have not issue in CV/LB gap so far, always within +/- 0.003</p>",
      "rawMarkdown": "I have not issue in CV/LB gap so far, always within +/- 0.003",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 649423,
      "author_name": "igormunizims",
      "author_url": "",
      "post_date": "10/15/2019 10:33:36",
      "content": "<p>I fix this by not applying smooth to my dice coef metric.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 659342,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "10/27/2019 13:42:33",
      "content": "<p>Kaggle computes dice for each pair of true and predicted and then averages them all. If you use the popular validation dice coef code in public notebooks, it doesn't do that. The correct metric to compile into your model is posted <a href=\"https://www.kaggle.com/c/understanding_cloud_organization/discussion/114093\">here</a>. Use the metric named <code>mean_dice_coef</code> in your models. (It uses the other function <code>single_dice_coef</code> but you don't use that one).</p>\n\n<p>There are other reasons why CV and LB may not correlate, but the above makes your model use the same scoring mechanism that Kaggle does.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 659413,
      "author_name": "naivelamb",
      "author_url": "",
      "post_date": "10/27/2019 16:10:00",
      "content": "<p>I have not issue in CV/LB gap so far, always within +/- 0.003</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "649420": "Hello. I see some kernels, and in many of them val_dice_coef is ~0.55, but LB score is 0.645.\n\nHow i can see my local score, which will be corelation with LB score.",
    "649423": "I fix this by not applying smooth to my dice coef metric.",
    "659342": "Kaggle computes dice for each pair of true and predicted and then averages them all. If you use the popular validation dice coef code in public notebooks, it doesn't do that. The correct metric to compile into your model is posted [here][1]. Use the metric named `mean_dice_coef` in your models. (It uses the other function `single_dice_coef` but you don't use that one).\n\nThere are other reasons why CV and LB may not correlate, but the above makes your model use the same scoring mechanism that Kaggle does.\n\n[1]: https://www.kaggle.com/c/understanding_cloud_organization/discussion/114093",
    "659413": "I have not issue in CV/LB gap so far, always within +/- 0.003"
  },
  "source": "meta"
}