{
  "id": 254343,
  "title": "Has anyone ever used TTA as if it might be useful?",
  "url": "/competitions/seti-breakthrough-listen/discussion/254343",
  "author_name": "zhangeng",
  "post_date": "2021-07-21T10:29:39.045000",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Can improve about 0.002 or even higher scores…</p>",
  "messages": [
    {
      "id": 1399446,
      "postDate": "2021-07-25T10:02:34.900Z",
      "content": "<p>Me too.</p>\n<p>My model improve about 0.001 higher score.</p>",
      "rawMarkdown": "Me too.\n\nMy model improve about 0.001 higher score.",
      "votes": 1
    },
    {
      "id": 1398308,
      "postDate": "2021-07-24T03:14:56.523Z",
      "content": "<p>1).efficientnet_b0:  cv=0.86906 lb=0.760<br>\n2).tta: lb=0.761<br>\n3).2avg:lb=0.763<br>\nOther models are almost the same, with an average increase of 0.002-0.003<br>\nexample：densenet201 increase 0.002</p>",
      "rawMarkdown": "1).efficientnet_b0:  cv=0.86906 lb=0.760\n2).tta: lb=0.761\n3).2avg:lb=0.763\nOther models are almost the same, with an average increase of 0.002-0.003\nexample：densenet201 increase 0.002",
      "votes": 1,
      "replies": [
        {
          "id": 1398594,
          "postDate": "2021-07-24T10:05:19.243Z",
          "content": "<p>what kind of TTA are you using, and how slower when you apply TTA? Thanks.</p>",
          "rawMarkdown": "what kind of TTA are you using, and how slower when you apply TTA? Thanks."
        },
        {
          "id": 1399430,
          "postDate": "2021-07-25T09:46:55.947Z",
          "content": "<p><a href=\"https://www.kaggle.com/superchenhao\" target=\"_blank\">@superchenhao</a> <br>\nWhen using TTA, the prediction program is about 2-3 times that without TTA</p>\n<p>spatial transformations (eg. horizontal, vertical, rotation, etc.)</p>\n<p>For prediction only, use TTA</p>\n<p>For the first mock exam, the average increase is 0.002.</p>\n<p>But not for ensemble</p>",
          "rawMarkdown": "@superchenhao \nWhen using TTA, the prediction program is about 2-3 times that without TTA\n\nspatial transformations (eg. horizontal, vertical, rotation, etc.)\n\nFor prediction only, use TTA\n\nFor the first mock exam, the average increase is 0.002.\n\nBut not for ensemble"
        }
      ]
    },
    {
      "id": 1395573,
      "postDate": "2021-07-21T10:29:39.047Z",
      "content": "<p>Can improve about 0.002 or even higher scores…</p>",
      "rawMarkdown": "Can improve about 0.002 or even higher scores...",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 1399446,
      "author_name": "mizoo",
      "author_url": "",
      "post_date": "2021-07-25T10:02:34.900000",
      "content": "<p>Me too.</p>\n<p>My model improve about 0.001 higher score.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1398308,
      "author_name": "zhangeng",
      "author_url": "",
      "post_date": "2021-07-24T03:14:56.523000",
      "content": "<p>1).efficientnet_b0:  cv=0.86906 lb=0.760<br>\n2).tta: lb=0.761<br>\n3).2avg:lb=0.763<br>\nOther models are almost the same, with an average increase of 0.002-0.003<br>\nexample：densenet201 increase 0.002</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1398594,
          "author_name": "Hao",
          "author_url": "",
          "post_date": "2021-07-24T10:05:19.243000",
          "content": "<p>what kind of TTA are you using, and how slower when you apply TTA? Thanks.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1399430,
          "author_name": "zhangeng",
          "author_url": "",
          "post_date": "2021-07-25T09:46:55.947000",
          "content": "<p><a href=\"https://www.kaggle.com/superchenhao\" target=\"_blank\">@superchenhao</a> <br>\nWhen using TTA, the prediction program is about 2-3 times that without TTA</p>\n<p>spatial transformations (eg. horizontal, vertical, rotation, etc.)</p>\n<p>For prediction only, use TTA</p>\n<p>For the first mock exam, the average increase is 0.002.</p>\n<p>But not for ensemble</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1399446": "Me too.\n\nMy model improve about 0.001 higher score.",
    "1398308": "1).efficientnet_b0:  cv=0.86906 lb=0.760\n2).tta: lb=0.761\n3).2avg:lb=0.763\nOther models are almost the same, with an average increase of 0.002-0.003\nexample：densenet201 increase 0.002",
    "1395573": "Can improve about 0.002 or even higher scores..."
  }
}