{
  "id": 428406,
  "title": "1024 training  did not correlate with lb",
  "url": "/competitions/google-research-identify-contrails-reduce-global-warming/discussion/428406",
  "author_name": "HZM",
  "post_date": "2023-08-01T09:24:48.736000",
  "votes": 1,
  "comment_count": 7,
  "views": 0,
  "content": "<p>i am using 1024 x 1024 as input got cv 0.69+ but lb only got 0.68, had annyone have the same issues?</p>",
  "messages": [
    {
      "id": 2371297,
      "postDate": "2023-08-03T03:54:34.363Z",
      "content": "<p>In our experiment, a larger image size (256×256-&gt;512×512) greatly improves the validation score (average ~0.015), but only slightly improves the lb score (average ~0.003) in four different models. </p>",
      "rawMarkdown": "In our experiment, a larger image size (256×256->512×512) greatly improves the validation score (average ~0.015), but only slightly improves the lb score (average ~0.003) in four different models. \n",
      "votes": 1
    },
    {
      "id": 2368728,
      "postDate": "2023-08-01T10:49:19.350Z",
      "content": "<p>I think there will shake. My CV &amp; LB is weak correlated. Simply HFlip almost no change to CV but +0.01 to LB. btw, what's your cv strategy?</p>",
      "rawMarkdown": "I think there will shake. My CV & LB is weak correlated. Simply HFlip almost no change to CV but +0.01 to LB. btw, what's your cv strategy?",
      "votes": 1,
      "replies": [
        {
          "id": 2368756,
          "postDate": "2023-08-01T11:07:02.030Z",
          "content": "<p>just val data provided in competation</p>",
          "rawMarkdown": "just val data provided in competation",
          "replies": [
            {
              "id": 2368823,
              "postDate": "2023-08-01T11:49:45.263Z",
              "content": "<p>You should try going for folds, you might end up very disapointed otherwise.</p>",
              "rawMarkdown": "You should try going for folds, you might end up very disapointed otherwise.",
              "votes": 2
            },
            {
              "id": 2368843,
              "postDate": "2023-08-01T12:11:10.800Z",
              "content": "<p>thanks, already changed as fold</p>",
              "rawMarkdown": "thanks, already changed as fold",
              "votes": 1
            }
          ]
        }
      ]
    },
    {
      "id": 2371243,
      "postDate": "2023-08-03T02:46:01.103Z",
      "content": "<p>Meet similar problem,  CV/LB weak correlted after cv &gt; 0.69 . </p>",
      "rawMarkdown": "Meet similar problem,  CV/LB weak correlted after cv > 0.69 . ",
      "votes": 2,
      "replies": [
        {
          "id": 2375705,
          "postDate": "2023-08-05T20:41:28.673Z",
          "content": "<p>local validation dice? or cross validation dice? reach 0.69</p>",
          "rawMarkdown": "local validation dice? or cross validation dice? reach 0.69"
        }
      ]
    },
    {
      "id": 2368632,
      "postDate": "2023-08-01T09:24:48.737Z",
      "content": "<p>i am using 1024 x 1024 as input got cv 0.69+ but lb only got 0.68, had annyone have the same issues?</p>",
      "rawMarkdown": "i am using 1024 x 1024 as input got cv 0.69+ but lb only got 0.68, had annyone have the same issues?"
    }
  ],
  "comments": [
    {
      "id": 2371297,
      "author_name": "棒棒碎冰冰",
      "author_url": "",
      "post_date": "2023-08-03T03:54:34.363000",
      "content": "<p>In our experiment, a larger image size (256×256-&gt;512×512) greatly improves the validation score (average ~0.015), but only slightly improves the lb score (average ~0.003) in four different models. </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 2368728,
      "author_name": "william.wu",
      "author_url": "",
      "post_date": "2023-08-01T10:49:19.350000",
      "content": "<p>I think there will shake. My CV &amp; LB is weak correlated. Simply HFlip almost no change to CV but +0.01 to LB. btw, what's your cv strategy?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 2368756,
          "author_name": "HZM",
          "author_url": "",
          "post_date": "2023-08-01T11:07:02.030000",
          "content": "<p>just val data provided in competation</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2368823,
              "author_name": "JEANMPIA",
              "author_url": "",
              "post_date": "2023-08-01T11:49:45.263000",
              "content": "<p>You should try going for folds, you might end up very disapointed otherwise.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2368843,
              "author_name": "HZM",
              "author_url": "",
              "post_date": "2023-08-01T12:11:10.800000",
              "content": "<p>thanks, already changed as fold</p>",
              "votes": 1,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2371243,
      "author_name": "KKY",
      "author_url": "",
      "post_date": "2023-08-03T02:46:01.103000",
      "content": "<p>Meet similar problem,  CV/LB weak correlted after cv &gt; 0.69 . </p>",
      "votes": 2,
      "replies": [
        {
          "id": 2375705,
          "author_name": "pky",
          "author_url": "",
          "post_date": "2023-08-05T20:41:28.673000",
          "content": "<p>local validation dice? or cross validation dice? reach 0.69</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2371297": "In our experiment, a larger image size (256×256->512×512) greatly improves the validation score (average ~0.015), but only slightly improves the lb score (average ~0.003) in four different models. \n",
    "2368728": "I think there will shake. My CV & LB is weak correlated. Simply HFlip almost no change to CV but +0.01 to LB. btw, what's your cv strategy?",
    "2371243": "Meet similar problem,  CV/LB weak correlted after cv > 0.69 . ",
    "2368632": "i am using 1024 x 1024 as input got cv 0.69+ but lb only got 0.68, had annyone have the same issues?"
  }
}