{
  "id": 134612,
  "title": "[HELP]: What is TTA and Label Smoothing? How to do it?",
  "url": "/competitions/bengaliai-cv19/discussion/134612",
  "author_name": "",
  "post_date": "2020-03-09T08:49:00.087888600Z",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I'm new at using TTA and label smoothing and heard many from many practitioner to talk about it. What are they actually and why we need them? How to implement it anyway?\nThank you.</p>",
  "messages": [
    {
      "id": "767162",
      "postDate": "03/09/2020 08:49:00",
      "content": "<p>I'm new at using TTA and label smoothing and heard many from many practitioner to talk about it. What are they actually and why we need them? How to implement it anyway?\nThank you.</p>",
      "rawMarkdown": "I'm new at using TTA and label smoothing and heard many from many practitioner to talk about it. What are they actually and why we need them? How to implement it anyway?\nThank you.",
      "votes": null
    },
    {
      "id": "767170",
      "postDate": "03/09/2020 09:01:06",
      "content": "<p>tta is test time augmentation.  idea is to use augmentations as you do in training, to create several variants of each image in test, predict on all variants, then combine the predictions for each image to get the final prediction.</p>\n\n<p>Label smoothing is used if you have a binary target, for instance if you one hot encode the target here.  Then you replace 0 by epsilon, and 1 by 1 - epsilon, where epsilon is a small value, say 0.1.  Someone shared code for it in the forum, search for it as I don't have the link handy.  In general reading all the forum is a good idea.</p>",
      "rawMarkdown": "tta is test time augmentation.  idea is to use augmentations as you do in training, to create several variants of each image in test, predict on all variants, then combine the predictions for each image to get the final prediction.\n\nLabel smoothing is used if you have a binary target, for instance if you one hot encode the target here.  Then you replace 0 by epsilon, and 1 by 1 - epsilon, where epsilon is a small value, say 0.1.  Someone shared code for it in the forum, search for it as I don't have the link handy.  In general reading all the forum is a good idea.",
      "votes": null
    },
    {
      "id": "767203",
      "postDate": "03/09/2020 09:52:57",
      "content": "<p>Thank you <a href=\"/cpmpml\">@cpmpml</a> </p>",
      "rawMarkdown": "Thank you @cpmpml",
      "votes": null
    },
    {
      "id": "767430",
      "postDate": "03/09/2020 16:09:27",
      "content": "<p>Here there is a nice post explaining label smoothing with examples using keras: <a href=\"https://www.pyimagesearch.com/2019/12/30/label-smoothing-with-keras-tensorflow-and-deep-learning/\">https://www.pyimagesearch.com/2019/12/30/label-smoothing-with-keras-tensorflow-and-deep-learning/</a></p>",
      "rawMarkdown": "Here there is a nice post explaining label smoothing with examples using keras: https://www.pyimagesearch.com/2019/12/30/label-smoothing-with-keras-tensorflow-and-deep-learning/",
      "votes": null
    },
    {
      "id": "767548",
      "postDate": "03/09/2020 19:41:17",
      "content": "<p>Thanks. \nLoved this comment</p>\n\n<blockquote>\n  <p>In general reading all the forum is a good idea. </p>\n</blockquote>",
      "rawMarkdown": "Thanks. \nLoved this comment\n&gt; In general reading all the forum is a good idea.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 767170,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "03/09/2020 09:01:06",
      "content": "<p>tta is test time augmentation.  idea is to use augmentations as you do in training, to create several variants of each image in test, predict on all variants, then combine the predictions for each image to get the final prediction.</p>\n\n<p>Label smoothing is used if you have a binary target, for instance if you one hot encode the target here.  Then you replace 0 by epsilon, and 1 by 1 - epsilon, where epsilon is a small value, say 0.1.  Someone shared code for it in the forum, search for it as I don't have the link handy.  In general reading all the forum is a good idea.</p>",
      "votes": null,
      "replies": [
        {
          "id": 767203,
          "author_name": "akashshingha850",
          "author_url": "",
          "post_date": "03/09/2020 09:52:57",
          "content": "<p>Thank you <a href=\"/cpmpml\">@cpmpml</a> </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 767548,
          "author_name": "kurianbenoy",
          "author_url": "",
          "post_date": "03/09/2020 19:41:17",
          "content": "<p>Thanks. \nLoved this comment</p>\n\n<blockquote>\n  <p>In general reading all the forum is a good idea. </p>\n</blockquote>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 767430,
      "author_name": "lbronchal",
      "author_url": "",
      "post_date": "03/09/2020 16:09:27",
      "content": "<p>Here there is a nice post explaining label smoothing with examples using keras: <a href=\"https://www.pyimagesearch.com/2019/12/30/label-smoothing-with-keras-tensorflow-and-deep-learning/\">https://www.pyimagesearch.com/2019/12/30/label-smoothing-with-keras-tensorflow-and-deep-learning/</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "767162": "I'm new at using TTA and label smoothing and heard many from many practitioner to talk about it. What are they actually and why we need them? How to implement it anyway?\nThank you.",
    "767170": "tta is test time augmentation.  idea is to use augmentations as you do in training, to create several variants of each image in test, predict on all variants, then combine the predictions for each image to get the final prediction.\n\nLabel smoothing is used if you have a binary target, for instance if you one hot encode the target here.  Then you replace 0 by epsilon, and 1 by 1 - epsilon, where epsilon is a small value, say 0.1.  Someone shared code for it in the forum, search for it as I don't have the link handy.  In general reading all the forum is a good idea.",
    "767203": "Thank you @cpmpml",
    "767430": "Here there is a nice post explaining label smoothing with examples using keras: https://www.pyimagesearch.com/2019/12/30/label-smoothing-with-keras-tensorflow-and-deep-learning/",
    "767548": "Thanks. \nLoved this comment\n&gt; In general reading all the forum is a good idea."
  },
  "source": "meta"
}