{
  "id": 251813,
  "title": "What TTA to use!?",
  "url": "/competitions/siim-covid19-detection/discussion/251813",
  "author_name": "",
  "post_date": "2021-07-09T01:08:10.605162Z",
  "votes": 2,
  "comment_count": 9,
  "views": 0,
  "content": "<p>I will be posting my before and after TTA results soon here and would like you to share yours too.</p>\n<p>Without TTA: .456<br>\nWith Hflip TTA: .457</p>",
  "messages": [
    {
      "id": "1381396",
      "postDate": "07/09/2021 01:08:10",
      "content": "<p>I will be posting my before and after TTA results soon here and would like you to share yours too.</p>\n<p>Without TTA: .456<br>\nWith Hflip TTA: .457</p>",
      "rawMarkdown": "I will be posting my before and after TTA results soon here and would like you to share yours too.\n\nWithout TTA: .456\nWith Hflip TTA: .457",
      "votes": null
    },
    {
      "id": "1381934",
      "postDate": "07/09/2021 11:11:53",
      "content": "<p>it all depends on the model. an increase from 0.001 to 0.009.</p>",
      "rawMarkdown": "it all depends on the model. an increase from 0.001 to 0.009.",
      "votes": null
    },
    {
      "id": "1381953",
      "postDate": "07/09/2021 11:39:30",
      "content": "<p>I've found a good use of TTA but not the same for scale. Usually, HFilp + RandomResizedCrop gives me a good boost but can't make scale work at all.</p>",
      "rawMarkdown": "I've found a good use of TTA but not the same for scale. Usually, HFilp + RandomResizedCrop gives me a good boost but can't make scale work at all.",
      "votes": null
    },
    {
      "id": "1382035",
      "postDate": "07/09/2021 12:57:05",
      "content": "<p>Omg, don't point it out.</p>",
      "rawMarkdown": "Omg, don't point it out.",
      "votes": null
    },
    {
      "id": "1382535",
      "postDate": "07/10/2021 02:27:23",
      "content": "<p>scaling image sizes works with EffV2</p>",
      "rawMarkdown": "scaling image sizes works with EffV2",
      "votes": null
    },
    {
      "id": "1382585",
      "postDate": "07/10/2021 04:01:34",
      "content": "<p>Did not work for me :-( … But to be fair I already have a really high score without the scale or any kinda tta.</p>",
      "rawMarkdown": "Did not work for me :-( ... But to be fair I already have a really high score without the scale or any kinda tta.",
      "votes": null
    },
    {
      "id": "1382916",
      "postDate": "07/10/2021 10:20:27",
      "content": "<p><a href=\"https://www.kaggle.com/harshitsheoran\" target=\"_blank\">@harshitsheoran</a> changing scaling at test time dint work for me either i think after 45.6 there not much to gain with TTA maybe flip but that's all.</p>",
      "rawMarkdown": "harshitsheoran changing scaling at test time dint work for me either i think after 45.6 there not much to gain with TTA maybe flip but that's all.",
      "votes": null
    },
    {
      "id": "1382981",
      "postDate": "07/10/2021 11:45:44",
      "content": "<p>Hflip is the basic one, since testset is small, we may have a lots of choices like 16x for 33x.</p>",
      "rawMarkdown": "Hflip is the basic one, since testset is small, we may have a lots of choices like 16x for 33x.",
      "votes": null
    },
    {
      "id": "1383905",
      "postDate": "07/11/2021 10:14:37",
      "content": "<p>Scaling didn't work for me either with EfficientnetV2 or I might just be getting my scales wrong 😑 </p>",
      "rawMarkdown": "Scaling didn't work for me either with EfficientnetV2 or I might just be getting my scales wrong 😑",
      "votes": null
    },
    {
      "id": "1390691",
      "postDate": "07/17/2021 01:50:48",
      "content": "<p>16x for 33x, mean 16 TTA  epochs for any 33x image_size data??</p>",
      "rawMarkdown": "16x for 33x, mean 16 TTA  epochs for any 33x image_size data??",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1381934,
      "author_name": "zaakciiru",
      "author_url": "",
      "post_date": "07/09/2021 11:11:53",
      "content": "<p>it all depends on the model. an increase from 0.001 to 0.009.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1381953,
      "author_name": "igormunizims",
      "author_url": "",
      "post_date": "07/09/2021 11:39:30",
      "content": "<p>I've found a good use of TTA but not the same for scale. Usually, HFilp + RandomResizedCrop gives me a good boost but can't make scale work at all.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1382535,
          "author_name": "varundutt9213",
          "author_url": "",
          "post_date": "07/10/2021 02:27:23",
          "content": "<p>scaling image sizes works with EffV2</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1382585,
          "author_name": "harshitsheoran",
          "author_url": "",
          "post_date": "07/10/2021 04:01:34",
          "content": "<p>Did not work for me :-( … But to be fair I already have a really high score without the scale or any kinda tta.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1382916,
          "author_name": "varundutt9213",
          "author_url": "",
          "post_date": "07/10/2021 10:20:27",
          "content": "<p><a href=\"https://www.kaggle.com/harshitsheoran\" target=\"_blank\">@harshitsheoran</a> changing scaling at test time dint work for me either i think after 45.6 there not much to gain with TTA maybe flip but that's all.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1382035,
      "author_name": "yus002",
      "author_url": "",
      "post_date": "07/09/2021 12:57:05",
      "content": "<p>Omg, don't point it out.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1382981,
      "author_name": "steamedsheep",
      "author_url": "",
      "post_date": "07/10/2021 11:45:44",
      "content": "<p>Hflip is the basic one, since testset is small, we may have a lots of choices like 16x for 33x.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1390691,
          "author_name": "vatsalmavani",
          "author_url": "",
          "post_date": "07/17/2021 01:50:48",
          "content": "<p>16x for 33x, mean 16 TTA  epochs for any 33x image_size data??</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1383905,
      "author_name": "benihime91",
      "author_url": "",
      "post_date": "07/11/2021 10:14:37",
      "content": "<p>Scaling didn't work for me either with EfficientnetV2 or I might just be getting my scales wrong 😑 </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1381396": "I will be posting my before and after TTA results soon here and would like you to share yours too.\n\nWithout TTA: .456\nWith Hflip TTA: .457",
    "1381934": "it all depends on the model. an increase from 0.001 to 0.009.",
    "1381953": "I've found a good use of TTA but not the same for scale. Usually, HFilp + RandomResizedCrop gives me a good boost but can't make scale work at all.",
    "1382035": "Omg, don't point it out.",
    "1382535": "scaling image sizes works with EffV2",
    "1382585": "Did not work for me :-( ... But to be fair I already have a really high score without the scale or any kinda tta.",
    "1382916": "harshitsheoran changing scaling at test time dint work for me either i think after 45.6 there not much to gain with TTA maybe flip but that's all.",
    "1382981": "Hflip is the basic one, since testset is small, we may have a lots of choices like 16x for 33x.",
    "1383905": "Scaling didn't work for me either with EfficientnetV2 or I might just be getting my scales wrong 😑",
    "1390691": "16x for 33x, mean 16 TTA  epochs for any 33x image_size data??"
  },
  "source": "meta"
}