{
  "id": 174199,
  "title": "Normalization - CoarseDropout",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/174199",
  "author_name": "",
  "post_date": "2020-08-12T15:57:56.713065300Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I am wondering for normalization if this should be performed before or after we use an augmentation technique like coarse dropout that zeroes out the cells.</p>\n<p>On one hand, I imagine that it makes sense to perform it first because zeroing out sections of the image will have a large effect on how the image will be recolored. On the other hand, I don't know if its more important that the aggregate statistics match imagenet.</p>\n<p>I feel like I don't know the real purpose of the normalization. Is it to make sure that the aggregate statistics of the image match imagenet? Or is to to make sure that the coloring of the image matches imagenet?</p>\n<p>How should I think about it?</p>",
  "messages": [
    {
      "id": "967974",
      "postDate": "08/12/2020 15:57:56",
      "content": "<p>I am wondering for normalization if this should be performed before or after we use an augmentation technique like coarse dropout that zeroes out the cells.</p>\n<p>On one hand, I imagine that it makes sense to perform it first because zeroing out sections of the image will have a large effect on how the image will be recolored. On the other hand, I don't know if its more important that the aggregate statistics match imagenet.</p>\n<p>I feel like I don't know the real purpose of the normalization. Is it to make sure that the aggregate statistics of the image match imagenet? Or is to to make sure that the coloring of the image matches imagenet?</p>\n<p>How should I think about it?</p>",
      "rawMarkdown": "I am wondering for normalization if this should be performed before or after we use an augmentation technique like coarse dropout that zeroes out the cells.\n\nOn one hand, I imagine that it makes sense to perform it first because zeroing out sections of the image will have a large effect on how the image will be recolored. On the other hand, I don't know if its more important that the aggregate statistics match imagenet.\n\nI feel like I don't know the real purpose of the normalization. Is it to make sure that the aggregate statistics of the image match imagenet? Or is to to make sure that the coloring of the image matches imagenet?\n\nHow should I think about it?",
      "votes": null
    },
    {
      "id": "968212",
      "postDate": "08/12/2020 19:16:30",
      "content": "<p>If I am not wrong Normalization is used to scale the pixel values (intensity). Normalization has a positive impact on convergence of loss and calculation of gradient. Additionally, checkout this answer on <a href=\"https://stats.stackexchange.com/questions/211436/why-normalize-images-by-subtracting-datasets-image-mean-instead-of-the-current\" target=\"_blank\">stackexchange</a> . Finally to answer your question, we perform normalization before performing transformations. </p>",
      "rawMarkdown": "If I am not wrong Normalization is used to scale the pixel values (intensity). Normalization has a positive impact on convergence of loss and calculation of gradient. Additionally, checkout this answer on [stackexchange](https://stats.stackexchange.com/questions/211436/why-normalize-images-by-subtracting-datasets-image-mean-instead-of-the-current) . Finally to answer your question, we perform normalization before performing transformations.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 968212,
      "author_name": "realsid",
      "author_url": "",
      "post_date": "08/12/2020 19:16:30",
      "content": "<p>If I am not wrong Normalization is used to scale the pixel values (intensity). Normalization has a positive impact on convergence of loss and calculation of gradient. Additionally, checkout this answer on <a href=\"https://stats.stackexchange.com/questions/211436/why-normalize-images-by-subtracting-datasets-image-mean-instead-of-the-current\" target=\"_blank\">stackexchange</a> . Finally to answer your question, we perform normalization before performing transformations. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "967974": "I am wondering for normalization if this should be performed before or after we use an augmentation technique like coarse dropout that zeroes out the cells.\n\nOn one hand, I imagine that it makes sense to perform it first because zeroing out sections of the image will have a large effect on how the image will be recolored. On the other hand, I don't know if its more important that the aggregate statistics match imagenet.\n\nI feel like I don't know the real purpose of the normalization. Is it to make sure that the aggregate statistics of the image match imagenet? Or is to to make sure that the coloring of the image matches imagenet?\n\nHow should I think about it?",
    "968212": "If I am not wrong Normalization is used to scale the pixel values (intensity). Normalization has a positive impact on convergence of loss and calculation of gradient. Additionally, checkout this answer on [stackexchange](https://stats.stackexchange.com/questions/211436/why-normalize-images-by-subtracting-datasets-image-mean-instead-of-the-current) . Finally to answer your question, we perform normalization before performing transformations."
  },
  "source": "meta"
}