{
  "id": 400617,
  "title": "is RMSE a good loss to compare models on?",
  "url": "/competitions/denoising-shabby-pages/discussion/400617",
  "author_name": "",
  "post_date": "2023-04-09T11:44:37.177390900Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I'd argue that rmse is not a good loss to compare the models. The gray levels of text seem like they can be arbitrarily different than the cleaned images. </p>",
  "messages": [
    {
      "id": "2215540",
      "postDate": "04/09/2023 11:44:37",
      "content": "<p>I'd argue that rmse is not a good loss to compare the models. The gray levels of text seem like they can be arbitrarily different than the cleaned images. </p>",
      "rawMarkdown": "I'd argue that rmse is not a good loss to compare the models. The gray levels of text seem like they can be arbitrarily different than the cleaned images.",
      "votes": null
    },
    {
      "id": "2215559",
      "postDate": "04/09/2023 12:08:56",
      "content": "<p>I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.</p>\n<p>In fact all my model use L1 , but the results are not as good as expected.</p>",
      "rawMarkdown": "I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.\n\nIn fact all my model use L1 , but the results are not as good as expected.",
      "votes": null
    },
    {
      "id": "2222933",
      "postDate": "04/15/2023 16:51:07",
      "content": "<blockquote>\n  <p>I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.</p>\n  <p>In fact all my model use L1 , but the results are not as good as expected.</p>\n</blockquote>\n<p>The reason I saw RMSE or even L1 is not a good loss is that the gray levels for text in the shabby and clean images are arbitrarily different. For example, in the shabby image the text could have a gray level of 20, while in the clean image, it could have a gray level of 30, why should one be correct and not the other? A perceptual loss would be a better fit for the task</p>",
      "rawMarkdown": "> I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.\n> \n> In fact all my model use L1 , but the results are not as good as expected.\n\nThe reason I saw RMSE or even L1 is not a good loss is that the gray levels for text in the shabby and clean images are arbitrarily different. For example, in the shabby image the text could have a gray level of 20, while in the clean image, it could have a gray level of 30, why should one be correct and not the other? A perceptual loss would be a better fit for the task",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2215559,
      "author_name": "qinmia",
      "author_url": "",
      "post_date": "04/09/2023 12:08:56",
      "content": "<p>I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.</p>\n<p>In fact all my model use L1 , but the results are not as good as expected.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2222933,
          "author_name": "cmaspi",
          "author_url": "",
          "post_date": "04/15/2023 16:51:07",
          "content": "<blockquote>\n  <p>I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.</p>\n  <p>In fact all my model use L1 , but the results are not as good as expected.</p>\n</blockquote>\n<p>The reason I saw RMSE or even L1 is not a good loss is that the gray levels for text in the shabby and clean images are arbitrarily different. For example, in the shabby image the text could have a gray level of 20, while in the clean image, it could have a gray level of 30, why should one be correct and not the other? A perceptual loss would be a better fit for the task</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2215540": "I'd argue that rmse is not a good loss to compare the models. The gray levels of text seem like they can be arbitrarily different than the cleaned images.",
    "2215559": "I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.\n\nIn fact all my model use L1 , but the results are not as good as expected.",
    "2222933": "> I think L1 may work a little better than RMSE, but it doesn't actually perform that well, so maybe there are other better options.\n> \n> In fact all my model use L1 , but the results are not as good as expected.\n\nThe reason I saw RMSE or even L1 is not a good loss is that the gray levels for text in the shabby and clean images are arbitrarily different. For example, in the shabby image the text could have a gray level of 20, while in the clean image, it could have a gray level of 30, why should one be correct and not the other? A perceptual loss would be a better fit for the task"
  },
  "source": "meta"
}