{
  "id": 241352,
  "title": "Is anyone ensembling and/or doing TTA yet?",
  "url": "/competitions/seti-breakthrough-listen/discussion/241352",
  "author_name": "James Howard",
  "post_date": "2021-05-24T07:49:44.205000",
  "votes": 12,
  "comment_count": 10,
  "views": 0,
  "content": "<p>I have yet to find any strategy that provides benefit above just using all 5 folds of my best model. TTA actually made this worse. This I found VERY surprising, because my models were trained with hflip and vflip. It makes me slightly nervous that I've accidentally overfit the LB…</p>",
  "messages": [
    {
      "id": 1320628,
      "postDate": "2021-05-24T07:49:44.207Z",
      "content": "<p>I have yet to find any strategy that provides benefit above just using all 5 folds of my best model. TTA actually made this worse. This I found VERY surprising, because my models were trained with hflip and vflip. It makes me slightly nervous that I've accidentally overfit the LB…</p>",
      "rawMarkdown": "I have yet to find any strategy that provides benefit above just using all 5 folds of my best model. TTA actually made this worse. This I found VERY surprising, because my models were trained with hflip and vflip. It makes me slightly nervous that I've accidentally overfit the LB...",
      "votes": 12
    },
    {
      "id": 1322228,
      "postDate": "2021-05-25T09:37:45.223Z",
      "content": "<p>TTA works for me.</p>",
      "rawMarkdown": "TTA works for me.",
      "votes": 5
    },
    {
      "id": 1321847,
      "postDate": "2021-05-25T03:23:48.843Z",
      "content": "<p>When using TTA I have often found it necessary to explore a pretty wide range of loops to determine where the values/benefit stabilized.    If you have the compute time to run experiment with 3 up to maybe 20 loops in your TTA you should find a point were results are consistent and show improvement over no TTA.  </p>",
      "rawMarkdown": "When using TTA I have often found it necessary to explore a pretty wide range of loops to determine where the values/benefit stabilized.    If you have the compute time to run experiment with 3 up to maybe 20 loops in your TTA you should find a point were results are consistent and show improvement over no TTA.  ",
      "votes": 3
    },
    {
      "id": 1330203,
      "postDate": "2021-05-31T16:30:30.740Z",
      "content": "<p>TTA definitely boosts the score in my case… But it is marginal… Yet to try ensemble…</p>",
      "rawMarkdown": "TTA definitely boosts the score in my case... But it is marginal... Yet to try ensemble...",
      "votes": 1
    },
    {
      "id": 1321805,
      "postDate": "2021-05-25T01:51:56.533Z",
      "content": "<p>i haven't  tried ensembling,but TTA work for me</p>",
      "rawMarkdown": "i haven't  tried ensembling,but TTA work for me",
      "votes": 1
    },
    {
      "id": 1322322,
      "postDate": "2021-05-25T11:18:22.493Z",
      "content": "<p>TTA do not work for me, ensemble does.</p>",
      "rawMarkdown": "TTA do not work for me, ensemble does.",
      "votes": 2
    },
    {
      "id": 1323111,
      "postDate": "2021-05-26T02:24:07.900Z",
      "content": "<p>Are you guys using TTA with albumentations or using <a href=\"https://github.com/qubvel/ttach\" target=\"_blank\">https://github.com/qubvel/ttach</a> ?</p>",
      "rawMarkdown": "Are you guys using TTA with albumentations or using https://github.com/qubvel/ttach ?",
      "replies": [
        {
          "id": 1329250,
          "postDate": "2021-05-31T02:21:29.027Z",
          "content": "<p>most of us use albumentations. I have never tried ttach. You can try both in same pipeline to compare results.</p>",
          "rawMarkdown": "most of us use albumentations. I have never tried ttach. You can try both in same pipeline to compare results."
        },
        {
          "id": 1329798,
          "postDate": "2021-05-31T11:30:22.843Z",
          "content": "<p>I've tried using albumentations for TTA but I got a worse result than no TTA.</p>",
          "rawMarkdown": "I've tried using albumentations for TTA but I got a worse result than no TTA."
        }
      ]
    },
    {
      "id": 1322582,
      "postDate": "2021-05-25T14:17:29.997Z",
      "content": "<p>I am a noob on this topic. So, what is TTA?</p>",
      "rawMarkdown": "I am a noob on this topic. So, what is TTA?",
      "replies": [
        {
          "id": 1322696,
          "postDate": "2021-05-25T15:49:25.840Z",
          "rawMarkdown": "",
          "votes": 4,
          "isDeleted": true
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1322228,
      "author_name": "Sergey Bryansky",
      "author_url": "",
      "post_date": "2021-05-25T09:37:45.223000",
      "content": "<p>TTA works for me.</p>",
      "votes": 5,
      "replies": []
    },
    {
      "id": 1321847,
      "author_name": "PC Jimmmy",
      "author_url": "",
      "post_date": "2021-05-25T03:23:48.843000",
      "content": "<p>When using TTA I have often found it necessary to explore a pretty wide range of loops to determine where the values/benefit stabilized.    If you have the compute time to run experiment with 3 up to maybe 20 loops in your TTA you should find a point were results are consistent and show improvement over no TTA.  </p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 1330203,
      "author_name": "Manav",
      "author_url": "",
      "post_date": "2021-05-31T16:30:30.740000",
      "content": "<p>TTA definitely boosts the score in my case… But it is marginal… Yet to try ensemble…</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1321805,
      "author_name": "FatLiuyun",
      "author_url": "",
      "post_date": "2021-05-25T01:51:56.533000",
      "content": "<p>i haven't  tried ensembling,but TTA work for me</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1322322,
      "author_name": "Chenglu",
      "author_url": "",
      "post_date": "2021-05-25T11:18:22.493000",
      "content": "<p>TTA do not work for me, ensemble does.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1323111,
      "author_name": "Eduardo Farina",
      "author_url": "",
      "post_date": "2021-05-26T02:24:07.900000",
      "content": "<p>Are you guys using TTA with albumentations or using <a href=\"https://github.com/qubvel/ttach\" target=\"_blank\">https://github.com/qubvel/ttach</a> ?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1329250,
          "author_name": "Aman Deep Gupta",
          "author_url": "",
          "post_date": "2021-05-31T02:21:29.027000",
          "content": "<p>most of us use albumentations. I have never tried ttach. You can try both in same pipeline to compare results.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1329798,
          "author_name": "Eduardo Farina",
          "author_url": "",
          "post_date": "2021-05-31T11:30:22.843000",
          "content": "<p>I've tried using albumentations for TTA but I got a worse result than no TTA.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1322582,
      "author_name": "Baran Hashemi",
      "author_url": "",
      "post_date": "2021-05-25T14:17:29.997000",
      "content": "<p>I am a noob on this topic. So, what is TTA?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1322696,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-05-25T15:49:25.840000",
          "content": "",
          "votes": 4,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1320628": "I have yet to find any strategy that provides benefit above just using all 5 folds of my best model. TTA actually made this worse. This I found VERY surprising, because my models were trained with hflip and vflip. It makes me slightly nervous that I've accidentally overfit the LB...",
    "1322228": "TTA works for me.",
    "1321847": "When using TTA I have often found it necessary to explore a pretty wide range of loops to determine where the values/benefit stabilized.    If you have the compute time to run experiment with 3 up to maybe 20 loops in your TTA you should find a point were results are consistent and show improvement over no TTA.  ",
    "1330203": "TTA definitely boosts the score in my case... But it is marginal... Yet to try ensemble...",
    "1321805": "i haven't  tried ensembling,but TTA work for me",
    "1322322": "TTA do not work for me, ensemble does.",
    "1323111": "Are you guys using TTA with albumentations or using https://github.com/qubvel/ttach ?",
    "1322582": "I am a noob on this topic. So, what is TTA?"
  }
}