{
  "id": 214236,
  "title": "Test Time Augmentations",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/214236",
  "author_name": "",
  "post_date": "2021-01-25T19:22:46.990642100Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi, I was wondering how test time augmentations should exactly work? If you use 3 test time augmentations, should these augmentations be set to be consistent? If so, how do you decide which ones to do?</p>",
  "messages": [
    {
      "id": "1169865",
      "postDate": "01/25/2021 19:22:46",
      "content": "<p>Hi, I was wondering how test time augmentations should exactly work? If you use 3 test time augmentations, should these augmentations be set to be consistent? If so, how do you decide which ones to do?</p>",
      "rawMarkdown": "Hi, I was wondering how test time augmentations should exactly work? If you use 3 test time augmentations, should these augmentations be set to be consistent? If so, how do you decide which ones to do?",
      "votes": null
    },
    {
      "id": "1171318",
      "postDate": "01/26/2021 19:17:06",
      "content": "<p>It is prefered to set them without any probabilities or you are not able to compare your scores with each other and every run could compute different scores.</p>\n<p>I usually take very basic augmentations for TTA:</p>\n<ul>\n<li>Flipping (Horizontal, Vertical)</li>\n<li>Rotation</li>\n</ul>",
      "rawMarkdown": "It is prefered to set them without any probabilities or you are not able to compare your scores with each other and every run could compute different scores.\n\nI usually take very basic augmentations for TTA:\n- Flipping (Horizontal, Vertical)\n- Rotation",
      "votes": null
    },
    {
      "id": "1171346",
      "postDate": "01/26/2021 19:26:45",
      "content": "<p>Would random cropping or center crop be a bad choice?</p>",
      "rawMarkdown": "Would random cropping or center crop be a bad choice?",
      "votes": null
    },
    {
      "id": "1171448",
      "postDate": "01/26/2021 20:34:15",
      "content": "<p>There is no must about augmentations to be constant. Like in the albumentation library, there is a class called <a href=\"https://albumentations.readthedocs.io/en/latest/probabilities.html\" target=\"_blank\">OneOf</a> which chooses from the list of augmentation techniques choice of yours with probability.</p>\n<p>Additional to that, there is a nice discussion <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/206724\" target=\"_blank\">here</a> which talks about the light vs hard augmentations specific to this competition.</p>\n<p>In my case light augmentations in the TTA increased the LB score</p>",
      "rawMarkdown": "There is no must about augmentations to be constant. Like in the albumentation library, there is a class called [OneOf](https://albumentations.readthedocs.io/en/latest/probabilities.html) which chooses from the list of augmentation techniques choice of yours with probability.\n\nAdditional to that, there is a nice discussion [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/206724) which talks about the light vs hard augmentations specific to this competition.\n\nIn my case light augmentations in the TTA increased the LB score",
      "votes": null
    },
    {
      "id": "1171597",
      "postDate": "01/27/2021 01:29:13",
      "content": "<p>same problem</p>",
      "rawMarkdown": "same problem",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1171318,
      "author_name": "aliabdin1",
      "author_url": "",
      "post_date": "01/26/2021 19:17:06",
      "content": "<p>It is prefered to set them without any probabilities or you are not able to compare your scores with each other and every run could compute different scores.</p>\n<p>I usually take very basic augmentations for TTA:</p>\n<ul>\n<li>Flipping (Horizontal, Vertical)</li>\n<li>Rotation</li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 1171346,
          "author_name": "ayu055",
          "author_url": "",
          "post_date": "01/26/2021 19:26:45",
          "content": "<p>Would random cropping or center crop be a bad choice?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1171597,
          "author_name": "zhangeng",
          "author_url": "",
          "post_date": "01/27/2021 01:29:13",
          "content": "<p>same problem</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1171448,
      "author_name": "snnclsr",
      "author_url": "",
      "post_date": "01/26/2021 20:34:15",
      "content": "<p>There is no must about augmentations to be constant. Like in the albumentation library, there is a class called <a href=\"https://albumentations.readthedocs.io/en/latest/probabilities.html\" target=\"_blank\">OneOf</a> which chooses from the list of augmentation techniques choice of yours with probability.</p>\n<p>Additional to that, there is a nice discussion <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/206724\" target=\"_blank\">here</a> which talks about the light vs hard augmentations specific to this competition.</p>\n<p>In my case light augmentations in the TTA increased the LB score</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1169865": "Hi, I was wondering how test time augmentations should exactly work? If you use 3 test time augmentations, should these augmentations be set to be consistent? If so, how do you decide which ones to do?",
    "1171318": "It is prefered to set them without any probabilities or you are not able to compare your scores with each other and every run could compute different scores.\n\nI usually take very basic augmentations for TTA:\n- Flipping (Horizontal, Vertical)\n- Rotation",
    "1171346": "Would random cropping or center crop be a bad choice?",
    "1171448": "There is no must about augmentations to be constant. Like in the albumentation library, there is a class called [OneOf](https://albumentations.readthedocs.io/en/latest/probabilities.html) which chooses from the list of augmentation techniques choice of yours with probability.\n\nAdditional to that, there is a nice discussion [here](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/206724) which talks about the light vs hard augmentations specific to this competition.\n\nIn my case light augmentations in the TTA increased the LB score",
    "1171597": "same problem"
  },
  "source": "meta"
}