{
  "id": 71353,
  "title": "Does TTA help?",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/71353",
  "author_name": "Zineng Tang",
  "post_date": "2018-11-13T01:00:19.118000",
  "votes": 5,
  "comment_count": 9,
  "views": 0,
  "content": "<p>Does left and right flip augmentation helps?</p>\n\n<p>I expect some methods on test time augmentation that work.</p>\n\n<p>Update:</p>\n\n<p>Now it looks like the improvement is not trivial.\nTTA with flip augmentation already might perform better.</p>",
  "messages": [
    {
      "id": 420039,
      "postDate": "2018-11-13T01:00:19.120Z",
      "content": "<p>Does left and right flip augmentation helps?</p>\n\n<p>I expect some methods on test time augmentation that work.</p>\n\n<p>Update:</p>\n\n<p>Now it looks like the improvement is not trivial.\nTTA with flip augmentation already might perform better.</p>",
      "rawMarkdown": "Does left and right flip augmentation helps?\n\nI expect some methods on test time augmentation that work.\n\n\n\nUpdate:\n\nNow it looks like the improvement is not trivial.\nTTA with flip augmentation already might perform better.",
      "votes": 5
    },
    {
      "id": 420536,
      "postDate": "2018-11-13T19:31:00.370Z",
      "content": "<p>For me it did help. I got +0.001 on the public LB by using horizontal flip. Also important to notice that the training of my LSTM was using augmentation, which also included horizontal flip.</p>",
      "rawMarkdown": "For me it did help. I got +0.001 on the public LB by using horizontal flip. Also important to notice that the training of my LSTM was using augmentation, which also included horizontal flip.",
      "votes": 3
    },
    {
      "id": 420608,
      "postDate": "2018-11-13T22:03:09.877Z",
      "content": "<p>horizontal flip +0.001 LB</p>",
      "rawMarkdown": "horizontal flip +0.001 LB",
      "votes": 2
    },
    {
      "id": 421354,
      "postDate": "2018-11-14T23:58:35.263Z",
      "content": "<p>For me, added hflip lower my LB 0.001</p>",
      "rawMarkdown": "For me, added hflip lower my LB 0.001"
    },
    {
      "id": 420835,
      "postDate": "2018-11-14T07:32:08.470Z",
      "content": "<p>Doesn't tried any TTA now, but hflip in training data augmentation make CV worse.</p>",
      "rawMarkdown": "Doesn't tried any TTA now, but hflip in training data augmentation make CV worse.",
      "replies": [
        {
          "id": 421185,
          "postDate": "2018-11-14T17:21:37.280Z",
          "content": "<p>Good to know, I haven't tried that.</p>",
          "rawMarkdown": "Good to know, I haven't tried that."
        }
      ]
    },
    {
      "id": 420724,
      "postDate": "2018-11-14T03:10:33.367Z",
      "content": "<p>I didn't use any augmentation, but I still do hflip TTA, and it gives ~0.006 improvement on the public LB.</p>\n\n<p>Tips: use weighted TTA, rather than a half-half averaging.</p>",
      "rawMarkdown": "I didn't use any augmentation, but I still do hflip TTA, and it gives ~0.006 improvement on the public LB.\n\nTips: use weighted TTA, rather than a half-half averaging.\n",
      "replies": [
        {
          "id": 420726,
          "postDate": "2018-11-14T03:13:23.833Z",
          "content": "<p>Hi, How did you assign the size of weigths to the model?</p>",
          "rawMarkdown": "Hi, How did you assign the size of weigths to the model?"
        },
        {
          "id": 420727,
          "postDate": "2018-11-14T03:18:54.887Z",
          "content": "<p>Hi @Gary-DeepLearning, I didn't get your question. What I meant is don't just simply average the two predictions (original and flipped).</p>\n\n<p>I am not 100% sure that I am always correct but I've tested it with the public 0.892 MobileNet kernel and I got 0.897 by just do the TTA without changing any other factor.</p>",
          "rawMarkdown": "Hi @Gary-DeepLearning, I didn't get your question. What I meant is don't just simply average the two predictions (original and flipped).\n\nI am not 100% sure that I am always correct but I've tested it with the public 0.892 MobileNet kernel and I got 0.897 by just do the TTA without changing any other factor."
        },
        {
          "id": 420757,
          "postDate": "2018-11-14T04:35:39.840Z",
          "content": "<p>Ok, thank you, I got it.</p>",
          "rawMarkdown": "Ok, thank you, I got it."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 420536,
      "author_name": "Nuno Ferreira",
      "author_url": "",
      "post_date": "2018-11-13T19:31:00.370000",
      "content": "<p>For me it did help. I got +0.001 on the public LB by using horizontal flip. Also important to notice that the training of my LSTM was using augmentation, which also included horizontal flip.</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 420608,
      "author_name": "Vladimir Nosov",
      "author_url": "",
      "post_date": "2018-11-13T22:03:09.877000",
      "content": "<p>horizontal flip +0.001 LB</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 421354,
      "author_name": "Strideradu",
      "author_url": "",
      "post_date": "2018-11-14T23:58:35.263000",
      "content": "<p>For me, added hflip lower my LB 0.001</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 420835,
      "author_name": "wh1te",
      "author_url": "",
      "post_date": "2018-11-14T07:32:08.470000",
      "content": "<p>Doesn't tried any TTA now, but hflip in training data augmentation make CV worse.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 421185,
          "author_name": "Kulbear",
          "author_url": "",
          "post_date": "2018-11-14T17:21:37.280000",
          "content": "<p>Good to know, I haven't tried that.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 420724,
      "author_name": "Kulbear",
      "author_url": "",
      "post_date": "2018-11-14T03:10:33.367000",
      "content": "<p>I didn't use any augmentation, but I still do hflip TTA, and it gives ~0.006 improvement on the public LB.</p>\n\n<p>Tips: use weighted TTA, rather than a half-half averaging.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 420726,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2018-11-14T03:13:23.833000",
          "content": "<p>Hi, How did you assign the size of weigths to the model?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 420727,
          "author_name": "Kulbear",
          "author_url": "",
          "post_date": "2018-11-14T03:18:54.887000",
          "content": "<p>Hi @Gary-DeepLearning, I didn't get your question. What I meant is don't just simply average the two predictions (original and flipped).</p>\n\n<p>I am not 100% sure that I am always correct but I've tested it with the public 0.892 MobileNet kernel and I got 0.897 by just do the TTA without changing any other factor.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 420757,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2018-11-14T04:35:39.840000",
          "content": "<p>Ok, thank you, I got it.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "420039": "Does left and right flip augmentation helps?\n\nI expect some methods on test time augmentation that work.\n\n\n\nUpdate:\n\nNow it looks like the improvement is not trivial.\nTTA with flip augmentation already might perform better.",
    "420536": "For me it did help. I got +0.001 on the public LB by using horizontal flip. Also important to notice that the training of my LSTM was using augmentation, which also included horizontal flip.",
    "420608": "horizontal flip +0.001 LB",
    "421354": "For me, added hflip lower my LB 0.001",
    "420835": "Doesn't tried any TTA now, but hflip in training data augmentation make CV worse.",
    "420724": "I didn't use any augmentation, but I still do hflip TTA, and it gives ~0.006 improvement on the public LB.\n\nTips: use weighted TTA, rather than a half-half averaging.\n"
  }
}