{
  "id": 295352,
  "title": "anyone tried TTA(Test Time Augmentation)?",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/295352",
  "author_name": "",
  "post_date": "2021-12-15T17:03:09.601344800Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>When inferring, TTA (Test Time Augmentation) gives a similar score or a lower score.<br>\nHas anyone tried TTA and got similar results?<br>\nAnd if TTA doesn't work in this competition, what could be the reason?</p>",
  "messages": [
    {
      "id": "1619143",
      "postDate": "12/15/2021 17:03:09",
      "content": "<p>When inferring, TTA (Test Time Augmentation) gives a similar score or a lower score.<br>\nHas anyone tried TTA and got similar results?<br>\nAnd if TTA doesn't work in this competition, what could be the reason?</p>",
      "rawMarkdown": "When inferring, TTA (Test Time Augmentation) gives a similar score or a lower score.\nHas anyone tried TTA and got similar results?\nAnd if TTA doesn't work in this competition, what could be the reason?",
      "votes": null
    },
    {
      "id": "1619357",
      "postDate": "12/15/2021 21:19:56",
      "content": "<p>Not yet. I am trying to find way for better model. But in future … propably yes.</p>",
      "rawMarkdown": "Not yet. I am trying to find way for better model. But in future ... propably yes.",
      "votes": null
    },
    {
      "id": "1619483",
      "postDate": "12/16/2021 01:50:33",
      "content": "<p>thank you.</p>\n<p>In general, I thought that TTA would work for object detection, but …<br>\nIs it unique to this competition? Further quest is needed.</p>",
      "rawMarkdown": "thank you.\n\nIn general, I thought that TTA would work for object detection, but ...\nIs it unique to this competition? Further quest is needed.",
      "votes": null
    },
    {
      "id": "1619553",
      "postDate": "12/16/2021 03:36:32",
      "content": "<p>I think TTA should work (it increases recall) since the comp metric is F2, recall is more important</p>",
      "rawMarkdown": "I think TTA should work (it increases recall) since the comp metric is F2, recall is more important",
      "votes": null
    },
    {
      "id": "1619853",
      "postDate": "12/16/2021 08:42:23",
      "content": "<p>Yes, we need to experiment. Since we have 2 months … we can experiment a lot. Now I have problem to build stable model with my monster 1x GPU …. 😂😂😂 Batch size 4-6 …. one epoch … 25 minutes …. 😂😂😂😪😭</p>",
      "rawMarkdown": "Yes, we need to experiment. Since we have 2 months ... we can experiment a lot. Now I have problem to build stable model with my monster 1x GPU .... 😂😂😂 Batch size 4-6 .... one epoch ... 25 minutes .... 😂😂😂😪😭",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1619357,
      "author_name": "remekkinas",
      "author_url": "",
      "post_date": "12/15/2021 21:19:56",
      "content": "<p>Not yet. I am trying to find way for better model. But in future … propably yes.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1619483,
          "author_name": "syurenuko",
          "author_url": "",
          "post_date": "12/16/2021 01:50:33",
          "content": "<p>thank you.</p>\n<p>In general, I thought that TTA would work for object detection, but …<br>\nIs it unique to this competition? Further quest is needed.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1619853,
          "author_name": "remekkinas",
          "author_url": "",
          "post_date": "12/16/2021 08:42:23",
          "content": "<p>Yes, we need to experiment. Since we have 2 months … we can experiment a lot. Now I have problem to build stable model with my monster 1x GPU …. 😂😂😂 Batch size 4-6 …. one epoch … 25 minutes …. 😂😂😂😪😭</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1619553,
      "author_name": "ptran1203",
      "author_url": "",
      "post_date": "12/16/2021 03:36:32",
      "content": "<p>I think TTA should work (it increases recall) since the comp metric is F2, recall is more important</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1619143": "When inferring, TTA (Test Time Augmentation) gives a similar score or a lower score.\nHas anyone tried TTA and got similar results?\nAnd if TTA doesn't work in this competition, what could be the reason?",
    "1619357": "Not yet. I am trying to find way for better model. But in future ... propably yes.",
    "1619483": "thank you.\n\nIn general, I thought that TTA would work for object detection, but ...\nIs it unique to this competition? Further quest is needed.",
    "1619553": "I think TTA should work (it increases recall) since the comp metric is F2, recall is more important",
    "1619853": "Yes, we need to experiment. Since we have 2 months ... we can experiment a lot. Now I have problem to build stable model with my monster 1x GPU .... 😂😂😂 Batch size 4-6 .... one epoch ... 25 minutes .... 😂😂😂😪😭"
  },
  "source": "meta"
}