{
  "id": 303861,
  "title": "Higher Image Resolution CV-LB Unstable Correlation",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/303861",
  "author_name": "",
  "post_date": "2022-01-29T18:24:04.595579400Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have noticed that using higher image resolutions for inferencing such as 3600 gives me a lower F2 score on my validation set (with the optimal confidence threshold), both with and without TTA. </p>\n<p>With TTA and image size 3600 , my CV F2 score is 0.532, whereas with TTA and image size 1280, my CV F2 score is 0.626.</p>\n<p>Without TTA and image size 3600, my CV F2 score is 0.282, whereas without TTA and image size 1280, my CV F2 score is 0.609.</p>\n<p>However, image size 3600 with TTA is still the one that performs the best on the public leaderboard, giving a score of 0.609 compared to 0.544 for image size 1280. Why is this happening and what does it mean?</p>",
  "messages": [
    {
      "id": "1668566",
      "postDate": "01/29/2022 18:24:04",
      "content": "<p>I have noticed that using higher image resolutions for inferencing such as 3600 gives me a lower F2 score on my validation set (with the optimal confidence threshold), both with and without TTA. </p>\n<p>With TTA and image size 3600 , my CV F2 score is 0.532, whereas with TTA and image size 1280, my CV F2 score is 0.626.</p>\n<p>Without TTA and image size 3600, my CV F2 score is 0.282, whereas without TTA and image size 1280, my CV F2 score is 0.609.</p>\n<p>However, image size 3600 with TTA is still the one that performs the best on the public leaderboard, giving a score of 0.609 compared to 0.544 for image size 1280. Why is this happening and what does it mean?</p>",
      "rawMarkdown": "I have noticed that using higher image resolutions for inferencing such as 3600 gives me a lower F2 score on my validation set (with the optimal confidence threshold), both with and without TTA. \n\nWith TTA and image size 3600 , my CV F2 score is 0.532, whereas with TTA and image size 1280, my CV F2 score is 0.626.\n\nWithout TTA and image size 3600, my CV F2 score is 0.282, whereas without TTA and image size 1280, my CV F2 score is 0.609.\n\nHowever, image size 3600 with TTA is still the one that performs the best on the public leaderboard, giving a score of 0.609 compared to 0.544 for image size 1280. Why is this happening and what does it mean?",
      "votes": null
    },
    {
      "id": "1672865",
      "postDate": "02/02/2022 10:57:31",
      "content": "<p>Hello. Please help novice.<br>\nLB - LeaderBoard.<br>\nCV - ?<br>\nTTA - ?</p>",
      "rawMarkdown": "Hello. Please help novice.\nLB - LeaderBoard.\nCV - ?\nTTA - ?",
      "votes": null
    },
    {
      "id": "1685120",
      "postDate": "02/11/2022 03:46:16",
      "content": "<p>Late but.. <br>\nCV - cross validation score<br>\nTTA - Test time augmentation</p>",
      "rawMarkdown": "Late but.. \nCV - cross validation score\nTTA - Test time augmentation",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1672865,
      "author_name": "antonsibilev",
      "author_url": "",
      "post_date": "02/02/2022 10:57:31",
      "content": "<p>Hello. Please help novice.<br>\nLB - LeaderBoard.<br>\nCV - ?<br>\nTTA - ?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1685120,
          "author_name": "learnitanyway",
          "author_url": "",
          "post_date": "02/11/2022 03:46:16",
          "content": "<p>Late but.. <br>\nCV - cross validation score<br>\nTTA - Test time augmentation</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1668566": "I have noticed that using higher image resolutions for inferencing such as 3600 gives me a lower F2 score on my validation set (with the optimal confidence threshold), both with and without TTA. \n\nWith TTA and image size 3600 , my CV F2 score is 0.532, whereas with TTA and image size 1280, my CV F2 score is 0.626.\n\nWithout TTA and image size 3600, my CV F2 score is 0.282, whereas without TTA and image size 1280, my CV F2 score is 0.609.\n\nHowever, image size 3600 with TTA is still the one that performs the best on the public leaderboard, giving a score of 0.609 compared to 0.544 for image size 1280. Why is this happening and what does it mean?",
    "1672865": "Hello. Please help novice.\nLB - LeaderBoard.\nCV - ?\nTTA - ?",
    "1685120": "Late but.. \nCV - cross validation score\nTTA - Test time augmentation"
  },
  "source": "meta"
}