{
  "id": 231920,
  "title": "Better loss, Worse LB",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/231920",
  "author_name": "Shihao Shao",
  "post_date": "2021-04-11T04:38:11.094000",
  "votes": 3,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I used BCELOSS and found that the when the validation loss decrease to some level, the LB score just have negative correlation with it.<br>\nDoes anyone face this issue?</p>",
  "messages": [
    {
      "id": 1269921,
      "postDate": "2021-04-11T04:38:11.093Z",
      "content": "<p>I used BCELOSS and found that the when the validation loss decrease to some level, the LB score just have negative correlation with it.<br>\nDoes anyone face this issue?</p>",
      "rawMarkdown": " I used BCELOSS and found that the when the validation loss decrease to some level, the LB score just have negative correlation with it.\n\nDoes anyone face this issue?",
      "votes": 3
    },
    {
      "id": 1270867,
      "postDate": "2021-04-12T05:37:08.060Z",
      "content": "<p>Validation is difficult in this challenge because we don't have any ground truth data at cell level. Good folds split can help, but it is not sufficient. I don't have any solution to offer though ;)</p>",
      "rawMarkdown": "Validation is difficult in this challenge because we don't have any ground truth data at cell level. Good folds split can help, but it is not sufficient. I don't have any solution to offer though ;)",
      "votes": 4,
      "replies": [
        {
          "id": 1271313,
          "postDate": "2021-04-12T13:24:32.420Z",
          "content": "<p>However, in slide level, I also meet this issue. We have the GT of the slide level data though..</p>",
          "rawMarkdown": "However, in slide level, I also meet this issue. We have the GT of the slide level data though.."
        }
      ]
    },
    {
      "id": 1269937,
      "postDate": "2021-04-11T05:26:11.100Z",
      "content": "<p>What split are you using for training and validation data? It maybe a really imbalanced one causing better loss to give worse LB.</p>",
      "rawMarkdown": "What split are you using for training and validation data? It maybe a really imbalanced one causing better loss to give worse LB.",
      "votes": 2,
      "replies": [
        {
          "id": 1269938,
          "postDate": "2021-04-11T05:30:32.450Z",
          "content": "<p>Just random split, which may be the cause.</p>",
          "rawMarkdown": "Just random split, which may be the cause."
        },
        {
          "id": 1269945,
          "postDate": "2021-04-11T05:38:01.057Z",
          "content": "<p>Try using a stratified split. It seems to be working well for this competition </p>",
          "rawMarkdown": "Try using a stratified split. It seems to be working well for this competition "
        },
        {
          "id": 1269947,
          "postDate": "2021-04-11T05:39:30.897Z",
          "content": "<p><a href=\"https://www.kaggle.com/p4rallax\" target=\"_blank\">@p4rallax</a> Thanks, I will try that.</p>",
          "rawMarkdown": "@p4rallax Thanks, I will try that."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1270867,
      "author_name": "Darek Kłeczek",
      "author_url": "",
      "post_date": "2021-04-12T05:37:08.060000",
      "content": "<p>Validation is difficult in this challenge because we don't have any ground truth data at cell level. Good folds split can help, but it is not sufficient. I don't have any solution to offer though ;)</p>",
      "votes": 4,
      "replies": [
        {
          "id": 1271313,
          "author_name": "Shihao Shao",
          "author_url": "",
          "post_date": "2021-04-12T13:24:32.420000",
          "content": "<p>However, in slide level, I also meet this issue. We have the GT of the slide level data though..</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1269937,
      "author_name": "Satwik",
      "author_url": "",
      "post_date": "2021-04-11T05:26:11.100000",
      "content": "<p>What split are you using for training and validation data? It maybe a really imbalanced one causing better loss to give worse LB.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1269938,
          "author_name": "Shihao Shao",
          "author_url": "",
          "post_date": "2021-04-11T05:30:32.450000",
          "content": "<p>Just random split, which may be the cause.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1269945,
          "author_name": "Satwik",
          "author_url": "",
          "post_date": "2021-04-11T05:38:01.057000",
          "content": "<p>Try using a stratified split. It seems to be working well for this competition </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1269947,
          "author_name": "Shihao Shao",
          "author_url": "",
          "post_date": "2021-04-11T05:39:30.897000",
          "content": "<p><a href=\"https://www.kaggle.com/p4rallax\" target=\"_blank\">@p4rallax</a> Thanks, I will try that.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1269921": " I used BCELOSS and found that the when the validation loss decrease to some level, the LB score just have negative correlation with it.\n\nDoes anyone face this issue?",
    "1270867": "Validation is difficult in this challenge because we don't have any ground truth data at cell level. Good folds split can help, but it is not sufficient. I don't have any solution to offer though ;)",
    "1269937": "What split are you using for training and validation data? It maybe a really imbalanced one causing better loss to give worse LB."
  }
}