{
  "id": 125221,
  "title": "Label Smoothing and TTA",
  "url": "/competitions/bengaliai-cv19/discussion/125221",
  "author_name": "",
  "post_date": "2020-01-09T10:25:50.008238900Z",
  "votes": 3,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Have anyone used both label smoothing technique and test time augmentation? Did you observe good results?</p>",
  "messages": [
    {
      "id": "714342",
      "postDate": "01/09/2020 10:25:50",
      "content": "<p>Have anyone used both label smoothing technique and test time augmentation? Did you observe good results?</p>",
      "rawMarkdown": "Have anyone used both label smoothing technique and test time augmentation? Did you observe good results?",
      "votes": null
    },
    {
      "id": "714605",
      "postDate": "01/09/2020 15:33:25",
      "content": "<p>Label smoothing showed minor improvements. Not Sure if TTA is a good idea. For the TTA you have to run 8 times model, which means we will exceed allowed GPU limit for this competition which is 2hours=) </p>\n\n<p>But again I might be wrong =) \nBest</p>",
      "rawMarkdown": "Label smoothing showed minor improvements. Not Sure if TTA is a good idea. For the TTA you have to run 8 times model, which means we will exceed allowed GPU limit for this competition which is 2hours=) \n\nBut again I might be wrong =) \nBest",
      "votes": null
    },
    {
      "id": "715000",
      "postDate": "01/10/2020 02:35:55",
      "content": "<p>Thanks DrHB.</p>",
      "rawMarkdown": "Thanks DrHB.",
      "votes": null
    },
    {
      "id": "715039",
      "postDate": "01/10/2020 04:13:18",
      "content": "<p>It's better to run 8 separate models than 8-fold TTA) especially since here there is no good TTA that contributes exactly the same as the main prediction (like flips or 90 degree rotations if the images had the corresponding symmetries).</p>",
      "rawMarkdown": "It's better to run 8 separate models than 8-fold TTA) especially since here there is no good TTA that contributes exactly the same as the main prediction (like flips or 90 degree rotations if the images had the corresponding symmetries).",
      "votes": null
    },
    {
      "id": "715366",
      "postDate": "01/10/2020 13:02:23",
      "content": "<p><a href=\"/iafoss\">@iafoss</a> We can use AutoAugment technique to find better methods. But i think it's computationally expensive. </p>",
      "rawMarkdown": "iafoss We can use AutoAugment technique to find better methods. But i think it's computationally expensive.",
      "votes": null
    },
    {
      "id": "766519",
      "postDate": "03/08/2020 09:38:44",
      "content": "<p>hi, i just wonder what kind of Augmentations are used in  your TTA, i just use Flip,Crop,but it did not work and even worse</p>",
      "rawMarkdown": "hi, i just wonder what kind of Augmentations are used in  your TTA, i just use Flip,Crop,but it did not work and even worse",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 714605,
      "author_name": "drhabib",
      "author_url": "",
      "post_date": "01/09/2020 15:33:25",
      "content": "<p>Label smoothing showed minor improvements. Not Sure if TTA is a good idea. For the TTA you have to run 8 times model, which means we will exceed allowed GPU limit for this competition which is 2hours=) </p>\n\n<p>But again I might be wrong =) \nBest</p>",
      "votes": null,
      "replies": [
        {
          "id": 715000,
          "author_name": "karan07",
          "author_url": "",
          "post_date": "01/10/2020 02:35:55",
          "content": "<p>Thanks DrHB.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 715039,
          "author_name": "iafoss",
          "author_url": "",
          "post_date": "01/10/2020 04:13:18",
          "content": "<p>It's better to run 8 separate models than 8-fold TTA) especially since here there is no good TTA that contributes exactly the same as the main prediction (like flips or 90 degree rotations if the images had the corresponding symmetries).</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 715366,
          "author_name": "karan07",
          "author_url": "",
          "post_date": "01/10/2020 13:02:23",
          "content": "<p><a href=\"/iafoss\">@iafoss</a> We can use AutoAugment technique to find better methods. But i think it's computationally expensive. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 766519,
      "author_name": "wuzexu",
      "author_url": "",
      "post_date": "03/08/2020 09:38:44",
      "content": "<p>hi, i just wonder what kind of Augmentations are used in  your TTA, i just use Flip,Crop,but it did not work and even worse</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "714342": "Have anyone used both label smoothing technique and test time augmentation? Did you observe good results?",
    "714605": "Label smoothing showed minor improvements. Not Sure if TTA is a good idea. For the TTA you have to run 8 times model, which means we will exceed allowed GPU limit for this competition which is 2hours=) \n\nBut again I might be wrong =) \nBest",
    "715000": "Thanks DrHB.",
    "715039": "It's better to run 8 separate models than 8-fold TTA) especially since here there is no good TTA that contributes exactly the same as the main prediction (like flips or 90 degree rotations if the images had the corresponding symmetries).",
    "715366": "iafoss We can use AutoAugment technique to find better methods. But i think it's computationally expensive.",
    "766519": "hi, i just wonder what kind of Augmentations are used in  your TTA, i just use Flip,Crop,but it did not work and even worse"
  },
  "source": "meta"
}