{
  "id": 202375,
  "title": "[Tips] How to use CoarseDropout with mask",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/202375",
  "author_name": "",
  "post_date": "2020-12-09T17:53:18.404542300Z",
  "votes": 7,
  "comment_count": 4,
  "views": 0,
  "content": "<h1>Intro</h1>\n<p>It is useful for those who want to apply various augmentation with mask.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F5cf25d216cb1c5d8e186e7e232c49d54%2F2.png?generation=1607536148590973&amp;alt=media\" alt=\"\"></p>\n<h1>CoarseDroput with mask</h1>\n<p>It can be easily applied using Albumentations.<br>\n<code>\nA.CoarseDropout(max_holes=8, max_height=20, max_width=20, mask_fill_value=0)\n</code><br>\n <img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F1fb70127b890e03c8633c5a6d55615e0%2F1.png?generation=1607535976485685&amp;alt=media\" alt=\"\"></p>\n<p>I Added more examples for augmentation in a public notebook.</p>\n<ul>\n<li>HorizontalFlip</li>\n<li>VerticalFlip</li>\n<li>RandomRotate90</li>\n<li>ShiftScaleRotate</li>\n<li>OpticalDistortion</li>\n<li>GridDistortion</li>\n<li>ElasticTransform</li>\n<li>CoarseDropout</li>\n</ul>\n<p>You can easily use these combinations.<br>\nPlease see more examples in <a href=\"https://www.kaggle.com/piantic/tutorial-augmentation-with-mask-visualizati-n?scriptVersionId=48925419\" target=\"_blank\">this notebook</a>.</p>\n<p>Thank you!</p>",
  "messages": [
    {
      "id": "1107464",
      "postDate": "12/09/2020 17:53:18",
      "content": "<h1>Intro</h1>\n<p>It is useful for those who want to apply various augmentation with mask.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F5cf25d216cb1c5d8e186e7e232c49d54%2F2.png?generation=1607536148590973&amp;alt=media\" alt=\"\"></p>\n<h1>CoarseDroput with mask</h1>\n<p>It can be easily applied using Albumentations.<br>\n<code>\nA.CoarseDropout(max_holes=8, max_height=20, max_width=20, mask_fill_value=0)\n</code><br>\n <img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F1fb70127b890e03c8633c5a6d55615e0%2F1.png?generation=1607535976485685&amp;alt=media\" alt=\"\"></p>\n<p>I Added more examples for augmentation in a public notebook.</p>\n<ul>\n<li>HorizontalFlip</li>\n<li>VerticalFlip</li>\n<li>RandomRotate90</li>\n<li>ShiftScaleRotate</li>\n<li>OpticalDistortion</li>\n<li>GridDistortion</li>\n<li>ElasticTransform</li>\n<li>CoarseDropout</li>\n</ul>\n<p>You can easily use these combinations.<br>\nPlease see more examples in <a href=\"https://www.kaggle.com/piantic/tutorial-augmentation-with-mask-visualizati-n?scriptVersionId=48925419\" target=\"_blank\">this notebook</a>.</p>\n<p>Thank you!</p>",
      "rawMarkdown": "# Intro\nIt is useful for those who want to apply various augmentation with mask.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F5cf25d216cb1c5d8e186e7e232c49d54%2F2.png?generation=1607536148590973&alt=media)\n\n# CoarseDroput with mask\nIt can be easily applied using Albumentations.\n`\nA.CoarseDropout(max_holes=8, max_height=20, max_width=20, mask_fill_value=0)\n`\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F1fb70127b890e03c8633c5a6d55615e0%2F1.png?generation=1607535976485685&alt=media)\n\nI Added more examples for augmentation in a public notebook.\n- HorizontalFlip\n- VerticalFlip\n- RandomRotate90\n- ShiftScaleRotate\n- OpticalDistortion\n- GridDistortion\n- ElasticTransform\n- CoarseDropout\n\nYou can easily use these combinations.\nPlease see more examples in [this notebook](https://www.kaggle.com/piantic/tutorial-augmentation-with-mask-visualizati-n?scriptVersionId=48925419).\n\nThank you!",
      "votes": null
    },
    {
      "id": "1108149",
      "postDate": "12/10/2020 10:39:43",
      "content": "<p>Does this improve the score?</p>",
      "rawMarkdown": "Does this improve the score?",
      "votes": null
    },
    {
      "id": "1108157",
      "postDate": "12/10/2020 10:48:34",
      "content": "<p>It still needs verification and It's hard to say that simply <code>CoarseDropout</code> improves lb score.<br>\nIf someone helps this, I would appreciate it if he or she can share it.</p>\n<p><a href=\"https://www.kaggle.com/zavodrobotov\" target=\"_blank\">@zavodrobotov</a> </p>",
      "rawMarkdown": "It still needs verification and It's hard to say that simply `CoarseDropout` improves lb score.\nIf someone helps this, I would appreciate it if he or she can share it.\n\n@zavodrobotov",
      "votes": null
    },
    {
      "id": "1108326",
      "postDate": "12/10/2020 14:46:43",
      "content": "<p>For us, it improves CV but hurts LB.</p>",
      "rawMarkdown": "For us, it improves CV but hurts LB.",
      "votes": null
    },
    {
      "id": "1118887",
      "postDate": "12/19/2020 13:50:45",
      "content": "<p>What about you  doing Dropout only with images, not mask?<br>\nIf there are holes only in image and not mask, model has to predict the hole's class by using around it.<br>\nThis could make more robust model because of more challenging data.</p>",
      "rawMarkdown": "What about you  doing Dropout only with images, not mask?\nIf there are holes only in image and not mask, model has to predict the hole's class by using around it.\nThis could make more robust model because of more challenging data.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1108149,
      "author_name": "zavodrobotov",
      "author_url": "",
      "post_date": "12/10/2020 10:39:43",
      "content": "<p>Does this improve the score?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1108157,
          "author_name": "piantic",
          "author_url": "",
          "post_date": "12/10/2020 10:48:34",
          "content": "<p>It still needs verification and It's hard to say that simply <code>CoarseDropout</code> improves lb score.<br>\nIf someone helps this, I would appreciate it if he or she can share it.</p>\n<p><a href=\"https://www.kaggle.com/zavodrobotov\" target=\"_blank\">@zavodrobotov</a> </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1108326,
          "author_name": "tugstugi",
          "author_url": "",
          "post_date": "12/10/2020 14:46:43",
          "content": "<p>For us, it improves CV but hurts LB.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1118887,
      "author_name": "curiosity806",
      "author_url": "",
      "post_date": "12/19/2020 13:50:45",
      "content": "<p>What about you  doing Dropout only with images, not mask?<br>\nIf there are holes only in image and not mask, model has to predict the hole's class by using around it.<br>\nThis could make more robust model because of more challenging data.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1107464": "# Intro\nIt is useful for those who want to apply various augmentation with mask.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F5cf25d216cb1c5d8e186e7e232c49d54%2F2.png?generation=1607536148590973&alt=media)\n\n# CoarseDroput with mask\nIt can be easily applied using Albumentations.\n`\nA.CoarseDropout(max_holes=8, max_height=20, max_width=20, mask_fill_value=0)\n`\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F1fb70127b890e03c8633c5a6d55615e0%2F1.png?generation=1607535976485685&alt=media)\n\nI Added more examples for augmentation in a public notebook.\n- HorizontalFlip\n- VerticalFlip\n- RandomRotate90\n- ShiftScaleRotate\n- OpticalDistortion\n- GridDistortion\n- ElasticTransform\n- CoarseDropout\n\nYou can easily use these combinations.\nPlease see more examples in [this notebook](https://www.kaggle.com/piantic/tutorial-augmentation-with-mask-visualizati-n?scriptVersionId=48925419).\n\nThank you!",
    "1108149": "Does this improve the score?",
    "1108157": "It still needs verification and It's hard to say that simply `CoarseDropout` improves lb score.\nIf someone helps this, I would appreciate it if he or she can share it.\n\n@zavodrobotov",
    "1108326": "For us, it improves CV but hurts LB.",
    "1118887": "What about you  doing Dropout only with images, not mask?\nIf there are holes only in image and not mask, model has to predict the hole's class by using around it.\nThis could make more robust model because of more challenging data."
  },
  "source": "meta"
}