{
  "id": 234578,
  "title": "TTA seems no improvements",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/234578",
  "author_name": "",
  "post_date": "2021-04-25T03:12:07.926344Z",
  "votes": 3,
  "comment_count": 9,
  "views": 0,
  "content": "<p>I used hflip, vf, hvf, but seems no obvious comtribution, compared with non tta. And I used the same augs during trainning. any other sharing?</p>",
  "messages": [
    {
      "id": "1283540",
      "postDate": "04/25/2021 03:12:07",
      "content": "<p>I used hflip, vf, hvf, but seems no obvious comtribution, compared with non tta. And I used the same augs during trainning. any other sharing?</p>",
      "rawMarkdown": "I used hflip, vf, hvf, but seems no obvious comtribution, compared with non tta. And I used the same augs during trainning. any other sharing?",
      "votes": null
    },
    {
      "id": "1283690",
      "postDate": "04/25/2021 07:22:22",
      "content": "<p>it is more stable and just believe its usage.</p>",
      "rawMarkdown": "it is more stable and just believe its usage.",
      "votes": null
    },
    {
      "id": "1283752",
      "postDate": "04/25/2021 08:24:01",
      "content": "<p>okay, thanks for comment :)</p>",
      "rawMarkdown": "okay, thanks for comment :)",
      "votes": null
    },
    {
      "id": "1283782",
      "postDate": "04/25/2021 09:03:05",
      "content": "<p>I find in general that ensembling and TTA isn't that impactful in this task. I think the networks do a fairly good job of being relatively spatially invariant in this task.</p>",
      "rawMarkdown": "I find in general that ensembling and TTA isn't that impactful in this task. I think the networks do a fairly good job of being relatively spatially invariant in this task.",
      "votes": null
    },
    {
      "id": "1283952",
      "postDate": "04/25/2021 12:39:56",
      "content": "<p>I face also the same behavior, I can't see improvement using rotate, vertical flip, horizontal flip TTA</p>",
      "rawMarkdown": "I face also the same behavior, I can't see improvement using rotate, vertical flip, horizontal flip TTA",
      "votes": null
    },
    {
      "id": "1285004",
      "postDate": "04/26/2021 13:43:29",
      "content": "<p>works for me. 0.001 improvement</p>",
      "rawMarkdown": "works for me. 0.001 improvement",
      "votes": null
    },
    {
      "id": "1285026",
      "postDate": "04/26/2021 14:04:24",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/andrasferenczi\" target=\"_blank\">@andrasferenczi</a> </p>\n<p>What TTA worked for you? </p>",
      "rawMarkdown": "Hi @andrasferenczi \n\nWhat TTA worked for you?",
      "votes": null
    },
    {
      "id": "1286095",
      "postDate": "04/27/2021 15:13:23",
      "content": "<p>I get about 0.012 in improvement going from no TTA to 8x augmentations on one specific model. 0.904-&gt; 0.916</p>",
      "rawMarkdown": "I get about 0.012 in improvement going from no TTA to 8x augmentations on one specific model. 0.904-> 0.916",
      "votes": null
    },
    {
      "id": "1286473",
      "postDate": "04/28/2021 03:40:41",
      "content": "<p>very impressive, I am confused about my 3xtta, why the results is so poor</p>",
      "rawMarkdown": "very impressive, I am confused about my 3xtta, why the results is so poor",
      "votes": null
    },
    {
      "id": "1286517",
      "postDate": "04/28/2021 04:58:41",
      "content": "<p>horizontal flip, vertical and 90 degree rotation</p>",
      "rawMarkdown": "horizontal flip, vertical and 90 degree rotation",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1283690,
      "author_name": "southsakura",
      "author_url": "",
      "post_date": "04/25/2021 07:22:22",
      "content": "<p>it is more stable and just believe its usage.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1283752,
          "author_name": "plugin1689",
          "author_url": "",
          "post_date": "04/25/2021 08:24:01",
          "content": "<p>okay, thanks for comment :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1283782,
      "author_name": "jamesphoward",
      "author_url": "",
      "post_date": "04/25/2021 09:03:05",
      "content": "<p>I find in general that ensembling and TTA isn't that impactful in this task. I think the networks do a fairly good job of being relatively spatially invariant in this task.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1283952,
      "author_name": "osamir",
      "author_url": "",
      "post_date": "04/25/2021 12:39:56",
      "content": "<p>I face also the same behavior, I can't see improvement using rotate, vertical flip, horizontal flip TTA</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1285004,
      "author_name": "andrasferenczi",
      "author_url": "",
      "post_date": "04/26/2021 13:43:29",
      "content": "<p>works for me. 0.001 improvement</p>",
      "votes": null,
      "replies": [
        {
          "id": 1285026,
          "author_name": "jagadish13",
          "author_url": "",
          "post_date": "04/26/2021 14:04:24",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/andrasferenczi\" target=\"_blank\">@andrasferenczi</a> </p>\n<p>What TTA worked for you? </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1286517,
          "author_name": "andrasferenczi",
          "author_url": "",
          "post_date": "04/28/2021 04:58:41",
          "content": "<p>horizontal flip, vertical and 90 degree rotation</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1286095,
      "author_name": "mistag",
      "author_url": "",
      "post_date": "04/27/2021 15:13:23",
      "content": "<p>I get about 0.012 in improvement going from no TTA to 8x augmentations on one specific model. 0.904-&gt; 0.916</p>",
      "votes": null,
      "replies": [
        {
          "id": 1286473,
          "author_name": "plugin1689",
          "author_url": "",
          "post_date": "04/28/2021 03:40:41",
          "content": "<p>very impressive, I am confused about my 3xtta, why the results is so poor</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1283540": "I used hflip, vf, hvf, but seems no obvious comtribution, compared with non tta. And I used the same augs during trainning. any other sharing?",
    "1283690": "it is more stable and just believe its usage.",
    "1283752": "okay, thanks for comment :)",
    "1283782": "I find in general that ensembling and TTA isn't that impactful in this task. I think the networks do a fairly good job of being relatively spatially invariant in this task.",
    "1283952": "I face also the same behavior, I can't see improvement using rotate, vertical flip, horizontal flip TTA",
    "1285004": "works for me. 0.001 improvement",
    "1285026": "Hi @andrasferenczi \n\nWhat TTA worked for you?",
    "1286095": "I get about 0.012 in improvement going from no TTA to 8x augmentations on one specific model. 0.904-> 0.916",
    "1286473": "very impressive, I am confused about my 3xtta, why the results is so poor",
    "1286517": "horizontal flip, vertical and 90 degree rotation"
  },
  "source": "meta"
}