{
  "id": 149747,
  "title": "Should I run albumentations on both train and valid?",
  "url": "/competitions/global-wheat-detection/discussion/149747",
  "author_name": "",
  "post_date": "2020-05-09T17:29:15.817032200Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have forked from this very useful notebook: <a href=\"https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train\">https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train</a> and have come across the Albumentations package which I hadn't seen before.</p>\n\n<p>I want to have a play around with the possible transforms but I notice in the original notebook the author has only included the flip transformation on the training set, not the validation set - through the get_train_transform/valid functions. </p>\n\n<p>Can someone tell me the reason for this, please? Should I also only include extra transformations on the training set and not validation?</p>\n\n<p>Thanks,</p>\n\n<p>BT</p>",
  "messages": [
    {
      "id": "839968",
      "postDate": "05/09/2020 17:29:15",
      "content": "<p>I have forked from this very useful notebook: <a href=\"https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train\">https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train</a> and have come across the Albumentations package which I hadn't seen before.</p>\n\n<p>I want to have a play around with the possible transforms but I notice in the original notebook the author has only included the flip transformation on the training set, not the validation set - through the get_train_transform/valid functions. </p>\n\n<p>Can someone tell me the reason for this, please? Should I also only include extra transformations on the training set and not validation?</p>\n\n<p>Thanks,</p>\n\n<p>BT</p>",
      "rawMarkdown": "I have forked from this very useful notebook: https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train and have come across the Albumentations package which I hadn't seen before.\n\nI want to have a play around with the possible transforms but I notice in the original notebook the author has only included the flip transformation on the training set, not the validation set - through the get_train_transform/valid functions. \n\nCan someone tell me the reason for this, please? Should I also only include extra transformations on the training set and not validation?\n\nThanks,\n\nBT",
      "votes": null
    },
    {
      "id": "840144",
      "postDate": "05/09/2020 19:43:25",
      "content": "<p>If you're just using standard transformations like rotations, flips, etc. I don't think there is any reason to run it on validation.</p>\n\n<p>The primary reason to use these is to artificially increase the dataset size and smooth the training distribution and reduce overfitting.</p>",
      "rawMarkdown": "If you're just using standard transformations like rotations, flips, etc. I don't think there is any reason to run it on validation.\n\nThe primary reason to use these is to artificially increase the dataset size and smooth the training distribution and reduce overfitting.",
      "votes": null
    },
    {
      "id": "860981",
      "postDate": "05/25/2020 18:37:10",
      "content": "<p>yes, that is call Test Time Augmentations</p>",
      "rawMarkdown": "yes, that is call Test Time Augmentations",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 840144,
      "author_name": "matthewmasters",
      "author_url": "",
      "post_date": "05/09/2020 19:43:25",
      "content": "<p>If you're just using standard transformations like rotations, flips, etc. I don't think there is any reason to run it on validation.</p>\n\n<p>The primary reason to use these is to artificially increase the dataset size and smooth the training distribution and reduce overfitting.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 860981,
      "author_name": "marcelosanchezortega",
      "author_url": "",
      "post_date": "05/25/2020 18:37:10",
      "content": "<p>yes, that is call Test Time Augmentations</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "839968": "I have forked from this very useful notebook: https://www.kaggle.com/pestipeti/pytorch-starter-fasterrcnn-train and have come across the Albumentations package which I hadn't seen before.\n\nI want to have a play around with the possible transforms but I notice in the original notebook the author has only included the flip transformation on the training set, not the validation set - through the get_train_transform/valid functions. \n\nCan someone tell me the reason for this, please? Should I also only include extra transformations on the training set and not validation?\n\nThanks,\n\nBT",
    "840144": "If you're just using standard transformations like rotations, flips, etc. I don't think there is any reason to run it on validation.\n\nThe primary reason to use these is to artificially increase the dataset size and smooth the training distribution and reduce overfitting.",
    "860981": "yes, that is call Test Time Augmentations"
  },
  "source": "meta"
}