{
  "id": 224984,
  "title": "Validation loss < training loss",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/224984",
  "author_name": "",
  "post_date": "2021-03-10T13:05:17.214280500Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>In some occasions <strong>validation loss</strong> &lt; <strong>training loss</strong>. Ever wondered why? </p>\n<p><img src=\"https://pbs.twimg.com/media/D2p-KsHU0AAjzVw?format=png&amp;name=small\" alt=\"photo\"></p>",
  "messages": [
    {
      "id": "1233480",
      "postDate": "03/10/2021 13:05:17",
      "content": "<p>In some occasions <strong>validation loss</strong> &lt; <strong>training loss</strong>. Ever wondered why? </p>\n<p><img src=\"https://pbs.twimg.com/media/D2p-KsHU0AAjzVw?format=png&amp;name=small\" alt=\"photo\"></p>",
      "rawMarkdown": "In some occasions **validation loss** < **training loss**. Ever wondered why? \n\n![photo](https://pbs.twimg.com/media/D2p-KsHU0AAjzVw?format=png&name=small)",
      "votes": null
    },
    {
      "id": "1233535",
      "postDate": "03/10/2021 13:55:11",
      "content": "<p>This can happen for a variety of reasons, but the most common one I can think of is that you apply complicated transformations like MixCut or CoarseDropOut to your training images, while you do not apply them during validation. So it's harder for the model to get good score on the training set while validation set is cleaner hence easier.</p>",
      "rawMarkdown": "This can happen for a variety of reasons, but the most common one I can think of is that you apply complicated transformations like MixCut or CoarseDropOut to your training images, while you do not apply them during validation. So it's harder for the model to get good score on the training set while validation set is cleaner hence easier.",
      "votes": null
    },
    {
      "id": "1233548",
      "postDate": "03/10/2021 13:56:52",
      "content": "<p>I would think the most common reason is regularization because it's applied during training, but not during validation &amp; testing</p>",
      "rawMarkdown": "I would think the most common reason is regularization because it's applied during training, but not during validation & testing",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1233535,
      "author_name": "optimo",
      "author_url": "",
      "post_date": "03/10/2021 13:55:11",
      "content": "<p>This can happen for a variety of reasons, but the most common one I can think of is that you apply complicated transformations like MixCut or CoarseDropOut to your training images, while you do not apply them during validation. So it's harder for the model to get good score on the training set while validation set is cleaner hence easier.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1233548,
          "author_name": "jolasa",
          "author_url": "",
          "post_date": "03/10/2021 13:56:52",
          "content": "<p>I would think the most common reason is regularization because it's applied during training, but not during validation &amp; testing</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1233480": "In some occasions **validation loss** < **training loss**. Ever wondered why? \n\n![photo](https://pbs.twimg.com/media/D2p-KsHU0AAjzVw?format=png&name=small)",
    "1233535": "This can happen for a variety of reasons, but the most common one I can think of is that you apply complicated transformations like MixCut or CoarseDropOut to your training images, while you do not apply them during validation. So it's harder for the model to get good score on the training set while validation set is cleaner hence easier.",
    "1233548": "I would think the most common reason is regularization because it's applied during training, but not during validation & testing"
  },
  "source": "meta"
}