{
  "id": 157435,
  "title": "Is it possible to use mixup in this competition? ",
  "url": "/competitions/alaska2-image-steganalysis/discussion/157435",
  "author_name": "",
  "post_date": "2020-06-10T17:20:05.033789700Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I am interested what will happen if we sum cover picture and the picture changed by any algorithm. Will it contain coded information? </p>",
  "messages": [
    {
      "id": "880981",
      "postDate": "06/10/2020 17:20:05",
      "content": "<p>I am interested what will happen if we sum cover picture and the picture changed by any algorithm. Will it contain coded information? </p>",
      "rawMarkdown": "I am interested what will happen if we sum cover picture and the picture changed by any algorithm. Will it contain coded information?",
      "votes": null
    },
    {
      "id": "893174",
      "postDate": "06/19/2020 12:40:55",
      "content": "<p>I don't think you can use mixup in this competition. It will change the pixel values of the target image. \nIf you sum cover picture and altered picture then you will have pixel values greater than 255. </p>\n\n<p>lm = np.random.beta(alpha, alpha)\ndata = lm*data + (1-lm)*other_data</p>\n\n<p>This a part of mixup. Weights of two different image should have sum equal to 1. \nI think, mixup may change the message inside a picture. Not sure if this will boost the model performance or not. <br>\nDo share if mixup helps you if you have tried it already...if you want </p>",
      "rawMarkdown": "I don't think you can use mixup in this competition. It will change the pixel values of the target image. \nIf you sum cover picture and altered picture then you will have pixel values greater than 255. \n\nlm = np.random.beta(alpha, alpha)\ndata = lm*data + (1-lm)*other_data\n\nThis a part of mixup. Weights of two different image should have sum equal to 1. \nI think, mixup may change the message inside a picture. Not sure if this will boost the model performance or not.  \nDo share if mixup helps you if you have tried it already...if you want",
      "votes": null
    },
    {
      "id": "900297",
      "postDate": "06/24/2020 18:12:21",
      "content": "<p>I was thinking about that as well. I see two possible outcomes:</p>\n\n<ul>\n<li>First, it completely corrupts the labels, hence a sharp drop in AUC score.\nOR</li>\n<li>Second, it is a useful augmentation technique in line with the task at hand and you can see an increase in score. </li>\n</ul>\n\n<p>Anyway, I feel like if you want to try Mixup, I would choose a beta close to 0.9, otherwise you might have completely unclassifiable images.</p>",
      "rawMarkdown": "I was thinking about that as well. I see two possible outcomes:\n\n- First, it completely corrupts the labels, hence a sharp drop in AUC score.\nOR\n- Second, it is a useful augmentation technique in line with the task at hand and you can see an increase in score. \n\nAnyway, I feel like if you want to try Mixup, I would choose a beta close to 0.9, otherwise you might have completely unclassifiable images.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 893174,
      "author_name": "zaber666",
      "author_url": "",
      "post_date": "06/19/2020 12:40:55",
      "content": "<p>I don't think you can use mixup in this competition. It will change the pixel values of the target image. \nIf you sum cover picture and altered picture then you will have pixel values greater than 255. </p>\n\n<p>lm = np.random.beta(alpha, alpha)\ndata = lm*data + (1-lm)*other_data</p>\n\n<p>This a part of mixup. Weights of two different image should have sum equal to 1. \nI think, mixup may change the message inside a picture. Not sure if this will boost the model performance or not. <br>\nDo share if mixup helps you if you have tried it already...if you want </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 900297,
      "author_name": "rftexas",
      "author_url": "",
      "post_date": "06/24/2020 18:12:21",
      "content": "<p>I was thinking about that as well. I see two possible outcomes:</p>\n\n<ul>\n<li>First, it completely corrupts the labels, hence a sharp drop in AUC score.\nOR</li>\n<li>Second, it is a useful augmentation technique in line with the task at hand and you can see an increase in score. </li>\n</ul>\n\n<p>Anyway, I feel like if you want to try Mixup, I would choose a beta close to 0.9, otherwise you might have completely unclassifiable images.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "880981": "I am interested what will happen if we sum cover picture and the picture changed by any algorithm. Will it contain coded information?",
    "893174": "I don't think you can use mixup in this competition. It will change the pixel values of the target image. \nIf you sum cover picture and altered picture then you will have pixel values greater than 255. \n\nlm = np.random.beta(alpha, alpha)\ndata = lm*data + (1-lm)*other_data\n\nThis a part of mixup. Weights of two different image should have sum equal to 1. \nI think, mixup may change the message inside a picture. Not sure if this will boost the model performance or not.  \nDo share if mixup helps you if you have tried it already...if you want",
    "900297": "I was thinking about that as well. I see two possible outcomes:\n\n- First, it completely corrupts the labels, hence a sharp drop in AUC score.\nOR\n- Second, it is a useful augmentation technique in line with the task at hand and you can see an increase in score. \n\nAnyway, I feel like if you want to try Mixup, I would choose a beta close to 0.9, otherwise you might have completely unclassifiable images."
  },
  "source": "meta"
}