{
  "id": 125223,
  "title": "Test time flip augumentation not working?",
  "url": "/competitions/pku-autonomous-driving/discussion/125223",
  "author_name": "",
  "post_date": "2020-01-09T10:35:00.383457800Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>During training, I apply mirror flips for data augumentation which benefits accuracy. So, mirror flips can be used for augumentation.</p>\n\n<p>However, trying test time flips decreases CV (simply adding the flipped-then-flipped-back and raw inference results). Have anyone tried digging into TTAs? Some discussion will be interesting.</p>",
  "messages": [
    {
      "id": "714349",
      "postDate": "01/09/2020 10:35:00",
      "content": "<p>During training, I apply mirror flips for data augumentation which benefits accuracy. So, mirror flips can be used for augumentation.</p>\n\n<p>However, trying test time flips decreases CV (simply adding the flipped-then-flipped-back and raw inference results). Have anyone tried digging into TTAs? Some discussion will be interesting.</p>",
      "rawMarkdown": "During training, I apply mirror flips for data augumentation which benefits accuracy. So, mirror flips can be used for augumentation.\n\nHowever, trying test time flips decreases CV (simply adding the flipped-then-flipped-back and raw inference results). Have anyone tried digging into TTAs? Some discussion will be interesting.",
      "votes": null
    },
    {
      "id": "714380",
      "postDate": "01/09/2020 11:23:39",
      "content": "<p>Yep, I tried it too (color, brightness and flip TTA). For me, it didn't work out on LB score and only lead to a very small gain in local validation score so I discarded that idea.</p>",
      "rawMarkdown": "Yep, I tried it too (color, brightness and flip TTA). For me, it didn't work out on LB score and only lead to a very small gain in local validation score so I discarded that idea.",
      "votes": null
    },
    {
      "id": "714437",
      "postDate": "01/09/2020 12:49:52",
      "content": "<p>Same as me. I tried horizontal flip TTA and it lead to slight decrease of both local cv and lb.\nI think this is because I set p=0.1 for horizontal flip augumentation when training. It might work if we increase the ratio.</p>",
      "rawMarkdown": "Same as me. I tried horizontal flip TTA and it lead to slight decrease of both local cv and lb.\nI think this is because I set p=0.1 for horizontal flip augumentation when training. It might work if we increase the ratio.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 714380,
      "author_name": "ilu000",
      "author_url": "",
      "post_date": "01/09/2020 11:23:39",
      "content": "<p>Yep, I tried it too (color, brightness and flip TTA). For me, it didn't work out on LB score and only lead to a very small gain in local validation score so I discarded that idea.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 714437,
      "author_name": "nihei123",
      "author_url": "",
      "post_date": "01/09/2020 12:49:52",
      "content": "<p>Same as me. I tried horizontal flip TTA and it lead to slight decrease of both local cv and lb.\nI think this is because I set p=0.1 for horizontal flip augumentation when training. It might work if we increase the ratio.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "714349": "During training, I apply mirror flips for data augumentation which benefits accuracy. So, mirror flips can be used for augumentation.\n\nHowever, trying test time flips decreases CV (simply adding the flipped-then-flipped-back and raw inference results). Have anyone tried digging into TTAs? Some discussion will be interesting.",
    "714380": "Yep, I tried it too (color, brightness and flip TTA). For me, it didn't work out on LB score and only lead to a very small gain in local validation score so I discarded that idea.",
    "714437": "Same as me. I tried horizontal flip TTA and it lead to slight decrease of both local cv and lb.\nI think this is because I set p=0.1 for horizontal flip augumentation when training. It might work if we increase the ratio."
  },
  "source": "meta"
}