{
  "id": 101944,
  "title": "Get Rid of TTA!!",
  "url": "/competitions/aptos2019-blindness-detection/discussion/101944",
  "author_name": "",
  "post_date": "2019-07-30T02:22:56.825314100Z",
  "votes": 6,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I am not sure how it will work with your model ,but what i realised is that TTA definitely \nhurt my  score ,I literally jumped from 492nd position to my current position in the leaderboard😉 .</p>",
  "messages": [
    {
      "id": "587991",
      "postDate": "07/30/2019 02:22:56",
      "content": "<p>I am not sure how it will work with your model ,but what i realised is that TTA definitely \nhurt my  score ,I literally jumped from 492nd position to my current position in the leaderboard😉 .</p>",
      "rawMarkdown": "I am not sure how it will work with your model ,but what i realised is that TTA definitely \nhurt my  score ,I literally jumped from 492nd position to my current position in the leaderboard😉 .",
      "votes": null
    },
    {
      "id": "588153",
      "postDate": "07/30/2019 08:17:15",
      "content": "<p>I agree!\nTTA improves CV score, but hurts LB score.</p>\n\n<p>Same for Cappa optimalization</p>",
      "rawMarkdown": "I agree!\nTTA improves CV score, but hurts LB score.\n\nSame for Cappa optimalization",
      "votes": null
    },
    {
      "id": "588198",
      "postDate": "07/30/2019 09:31:05",
      "content": "<p>All TTAs are not made equal. If you perform \"aggressive\" TTA, with a lot of transformations that deeply modify your initial image, then yes it could be detrimental. \"Soft\" TTA with small transformations should still improve the robustness of your prediction process</p>",
      "rawMarkdown": "All TTAs are not made equal. If you perform \"aggressive\" TTA, with a lot of transformations that deeply modify your initial image, then yes it could be detrimental. \"Soft\" TTA with small transformations should still improve the robustness of your prediction process",
      "votes": null
    },
    {
      "id": "588323",
      "postDate": "07/30/2019 12:24:59",
      "content": "<p>Ya well said .</p>",
      "rawMarkdown": "Ya well said .",
      "votes": null
    },
    {
      "id": "589209",
      "postDate": "07/31/2019 16:17:18",
      "content": "<p>Same here, it not only hurts the LB score but it also dramatically increases the commit&amp;submission times.\nI went from waiting ~2 hours to ~5 hours.</p>",
      "rawMarkdown": "Same here, it not only hurts the LB score but it also dramatically increases the commit&amp;submission times.\nI went from waiting ~2 hours to ~5 hours.",
      "votes": null
    },
    {
      "id": "591111",
      "postDate": "08/03/2019 07:01:12",
      "content": "<p>Thanks! Are you using ResNet?</p>",
      "rawMarkdown": "Thanks! Are you using ResNet?",
      "votes": null
    },
    {
      "id": "591120",
      "postDate": "08/03/2019 07:16:23",
      "content": "<p>Densenet is what i am using.</p>",
      "rawMarkdown": "Densenet is what i am using.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 588153,
      "author_name": "nemethpeti",
      "author_url": "",
      "post_date": "07/30/2019 08:17:15",
      "content": "<p>I agree!\nTTA improves CV score, but hurts LB score.</p>\n\n<p>Same for Cappa optimalization</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 588198,
      "author_name": "juliencs",
      "author_url": "",
      "post_date": "07/30/2019 09:31:05",
      "content": "<p>All TTAs are not made equal. If you perform \"aggressive\" TTA, with a lot of transformations that deeply modify your initial image, then yes it could be detrimental. \"Soft\" TTA with small transformations should still improve the robustness of your prediction process</p>",
      "votes": null,
      "replies": [
        {
          "id": 588323,
          "author_name": "amardeepganguly",
          "author_url": "",
          "post_date": "07/30/2019 12:24:59",
          "content": "<p>Ya well said .</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 589209,
      "author_name": "sam1320",
      "author_url": "",
      "post_date": "07/31/2019 16:17:18",
      "content": "<p>Same here, it not only hurts the LB score but it also dramatically increases the commit&amp;submission times.\nI went from waiting ~2 hours to ~5 hours.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 591111,
      "author_name": "kir486680",
      "author_url": "",
      "post_date": "08/03/2019 07:01:12",
      "content": "<p>Thanks! Are you using ResNet?</p>",
      "votes": null,
      "replies": [
        {
          "id": 591120,
          "author_name": "amardeepganguly",
          "author_url": "",
          "post_date": "08/03/2019 07:16:23",
          "content": "<p>Densenet is what i am using.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "587991": "I am not sure how it will work with your model ,but what i realised is that TTA definitely \nhurt my  score ,I literally jumped from 492nd position to my current position in the leaderboard😉 .",
    "588153": "I agree!\nTTA improves CV score, but hurts LB score.\n\nSame for Cappa optimalization",
    "588198": "All TTAs are not made equal. If you perform \"aggressive\" TTA, with a lot of transformations that deeply modify your initial image, then yes it could be detrimental. \"Soft\" TTA with small transformations should still improve the robustness of your prediction process",
    "588323": "Ya well said .",
    "589209": "Same here, it not only hurts the LB score but it also dramatically increases the commit&amp;submission times.\nI went from waiting ~2 hours to ~5 hours.",
    "591111": "Thanks! Are you using ResNet?",
    "591120": "Densenet is what i am using."
  },
  "source": "meta"
}