{
  "id": 177623,
  "title": " Good results with image TTA inference?",
  "url": "/competitions/birdsong-recognition/discussion/177623",
  "author_name": "",
  "post_date": "2020-08-26T17:04:39.458142600Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>This topic is a tip for those who have not yet tried TTA inference to save them time or I need advice along the way, one of them applies.</p>\n<p>I have tried some combinations with TTA inference, </p>\n<ul>\n<li>with models trained with TTA and without</li>\n<li>with ensemble and/or stacking</li>\n<li>random TTA and fixed</li>\n<li>horizental, vertical, brightness, contrast and also rotate for sameDim x sameDim images.</li>\n<li>etc<br>\nNothing works better than without TTA :/</li>\n</ul>\n<p>Maybe one should exclude it also for training? </p>",
  "messages": [
    {
      "id": "986655",
      "postDate": "08/26/2020 17:04:39",
      "content": "<p>This topic is a tip for those who have not yet tried TTA inference to save them time or I need advice along the way, one of them applies.</p>\n<p>I have tried some combinations with TTA inference, </p>\n<ul>\n<li>with models trained with TTA and without</li>\n<li>with ensemble and/or stacking</li>\n<li>random TTA and fixed</li>\n<li>horizental, vertical, brightness, contrast and also rotate for sameDim x sameDim images.</li>\n<li>etc<br>\nNothing works better than without TTA :/</li>\n</ul>\n<p>Maybe one should exclude it also for training? </p>",
      "rawMarkdown": "This topic is a tip for those who have not yet tried TTA inference to save them time or I need advice along the way, one of them applies.\n\nI have tried some combinations with TTA inference, \n- with models trained with TTA and without\n- with ensemble and/or stacking\n- random TTA and fixed\n- horizental, vertical, brightness, contrast and also rotate for sameDim x sameDim images.\n- etc\n Nothing works better than without TTA :/\n\nMaybe one should exclude it also for training?",
      "votes": null
    },
    {
      "id": "986686",
      "postDate": "08/26/2020 17:30:04",
      "content": "<p>Same here.</p>",
      "rawMarkdown": "Same here.",
      "votes": null
    },
    {
      "id": "987125",
      "postDate": "08/27/2020 03:25:15",
      "content": "<p>my LB decreased with flipping</p>",
      "rawMarkdown": "my LB decreased with flipping",
      "votes": null
    },
    {
      "id": "987355",
      "postDate": "08/27/2020 07:28:13",
      "content": "<p>The Spectrogram shows different features of frequencies. Its characteristic changed after flipped/mirrowed. So it cound't work.</p>",
      "rawMarkdown": "The Spectrogram shows different features of frequencies. Its characteristic changed after flipped/mirrowed. So it cound't work.",
      "votes": null
    },
    {
      "id": "987359",
      "postDate": "08/27/2020 07:34:21",
      "content": "<p>Ok thanks, when I view at the image before and after TTA visually(which is important) it looks like any usually TTA but maybe it's more changes than the eye can see. (my first competition with sound)</p>",
      "rawMarkdown": "Ok thanks, when I view at the image before and after TTA visually(which is important) it looks like any usually TTA but maybe it's more changes than the eye can see. (my first competition with sound)",
      "votes": null
    },
    {
      "id": "988270",
      "postDate": "08/27/2020 23:58:15",
      "content": "<p>I think flipping decreases LB because essentially the x-axis gives time-domain information. Flipping would give the same effect as if the sdrib era gnikaeps sdrawkcab 😜</p>",
      "rawMarkdown": "I think flipping decreases LB because essentially the x-axis gives time-domain information. Flipping would give the same effect as if the sdrib era gnikaeps sdrawkcab 😜",
      "votes": null
    },
    {
      "id": "988605",
      "postDate": "08/28/2020 06:39:37",
      "content": "<p>Oh no, the <strong>flip side of TTA</strong>,  it became a <strong>bi(u)rd-en!</strong><br>\n;)</p>\n<p>Have though read about TTA in previous audio top solutions, maybe one have to train longer with tta or a specific audio tta solution. </p>",
      "rawMarkdown": "Oh no, the **flip side of TTA**,  it became a **bi(u)rd-en!**\n;)\n\nHave though read about TTA in previous audio top solutions, maybe one have to train longer with tta or a specific audio tta solution.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 986686,
      "author_name": "mpware",
      "author_url": "",
      "post_date": "08/26/2020 17:30:04",
      "content": "<p>Same here.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 987355,
      "author_name": "gzl0506",
      "author_url": "",
      "post_date": "08/27/2020 07:28:13",
      "content": "<p>The Spectrogram shows different features of frequencies. Its characteristic changed after flipped/mirrowed. So it cound't work.</p>",
      "votes": null,
      "replies": [
        {
          "id": 987359,
          "author_name": "kirderf",
          "author_url": "",
          "post_date": "08/27/2020 07:34:21",
          "content": "<p>Ok thanks, when I view at the image before and after TTA visually(which is important) it looks like any usually TTA but maybe it's more changes than the eye can see. (my first competition with sound)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 988270,
      "author_name": "mightyrains",
      "author_url": "",
      "post_date": "08/27/2020 23:58:15",
      "content": "<p>I think flipping decreases LB because essentially the x-axis gives time-domain information. Flipping would give the same effect as if the sdrib era gnikaeps sdrawkcab 😜</p>",
      "votes": null,
      "replies": [
        {
          "id": 988605,
          "author_name": "kirderf",
          "author_url": "",
          "post_date": "08/28/2020 06:39:37",
          "content": "<p>Oh no, the <strong>flip side of TTA</strong>,  it became a <strong>bi(u)rd-en!</strong><br>\n;)</p>\n<p>Have though read about TTA in previous audio top solutions, maybe one have to train longer with tta or a specific audio tta solution. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 987125,
      "author_name": "yash612",
      "author_url": "",
      "post_date": "08/27/2020 03:25:15",
      "content": "<p>my LB decreased with flipping</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "986655": "This topic is a tip for those who have not yet tried TTA inference to save them time or I need advice along the way, one of them applies.\n\nI have tried some combinations with TTA inference, \n- with models trained with TTA and without\n- with ensemble and/or stacking\n- random TTA and fixed\n- horizental, vertical, brightness, contrast and also rotate for sameDim x sameDim images.\n- etc\n Nothing works better than without TTA :/\n\nMaybe one should exclude it also for training?",
    "986686": "Same here.",
    "987125": "my LB decreased with flipping",
    "987355": "The Spectrogram shows different features of frequencies. Its characteristic changed after flipped/mirrowed. So it cound't work.",
    "987359": "Ok thanks, when I view at the image before and after TTA visually(which is important) it looks like any usually TTA but maybe it's more changes than the eye can see. (my first competition with sound)",
    "988270": "I think flipping decreases LB because essentially the x-axis gives time-domain information. Flipping would give the same effect as if the sdrib era gnikaeps sdrawkcab 😜",
    "988605": "Oh no, the **flip side of TTA**,  it became a **bi(u)rd-en!**\n;)\n\nHave though read about TTA in previous audio top solutions, maybe one have to train longer with tta or a specific audio tta solution."
  },
  "source": "meta"
}