{
  "id": 140096,
  "title": "Why predictions of my model so skewed towards fakes?",
  "url": "/competitions/deepfake-detection-challenge/discussion/140096",
  "author_name": "",
  "post_date": "2020-03-31T09:50:22.487566400Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<ol>\n<li>I have filtered video files by actors</li>\n<li>Then I extracted faces from all frames from real videos and for each video pick (15 * number of relative fake vids) frames by clustering most different faces using KMeans</li>\n<li>Then I distributed these frames on relative fake vids and extract faces from them</li>\n</ol>\n\n<p>After training Resnet my predictions skewed to the right:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3920094%2F76be758b956cb4f2189ae767318df537%2Fdownload.png?generation=1585648129548914&amp;alt=media\" alt=\"\"></p>\n\n<p>Why is that? Is it because fake faces much more diverse than real faces? How to deal with that?</p>",
  "messages": [
    {
      "id": "792538",
      "postDate": "03/31/2020 09:50:22",
      "content": "<ol>\n<li>I have filtered video files by actors</li>\n<li>Then I extracted faces from all frames from real videos and for each video pick (15 * number of relative fake vids) frames by clustering most different faces using KMeans</li>\n<li>Then I distributed these frames on relative fake vids and extract faces from them</li>\n</ol>\n\n<p>After training Resnet my predictions skewed to the right:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3920094%2F76be758b956cb4f2189ae767318df537%2Fdownload.png?generation=1585648129548914&amp;alt=media\" alt=\"\"></p>\n\n<p>Why is that? Is it because fake faces much more diverse than real faces? How to deal with that?</p>",
      "rawMarkdown": "1. I have filtered video files by actors\n2. Then I extracted faces from all frames from real videos and for each video pick (15 * number of relative fake vids) frames by clustering most different faces using KMeans\n3. Then I distributed these frames on relative fake vids and extract faces from them\n\nAfter training Resnet my predictions skewed to the right:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3920094%2F76be758b956cb4f2189ae767318df537%2Fdownload.png?generation=1585648129548914&amp;alt=media)\n\nWhy is that? Is it because fake faces much more diverse than real faces? How to deal with that?",
      "votes": null
    },
    {
      "id": "792969",
      "postDate": "03/31/2020 17:28:58",
      "content": "<p>Did you undersample the fake samples? </p>",
      "rawMarkdown": "Did you undersample the fake samples?",
      "votes": null
    },
    {
      "id": "792972",
      "postDate": "03/31/2020 17:32:59",
      "content": "<p>number of fake faces = number of real faces</p>",
      "rawMarkdown": "number of fake faces = number of real faces",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 792969,
      "author_name": "vpaslay",
      "author_url": "",
      "post_date": "03/31/2020 17:28:58",
      "content": "<p>Did you undersample the fake samples? </p>",
      "votes": null,
      "replies": [
        {
          "id": 792972,
          "author_name": "mxmka87",
          "author_url": "",
          "post_date": "03/31/2020 17:32:59",
          "content": "<p>number of fake faces = number of real faces</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "792538": "1. I have filtered video files by actors\n2. Then I extracted faces from all frames from real videos and for each video pick (15 * number of relative fake vids) frames by clustering most different faces using KMeans\n3. Then I distributed these frames on relative fake vids and extract faces from them\n\nAfter training Resnet my predictions skewed to the right:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3920094%2F76be758b956cb4f2189ae767318df537%2Fdownload.png?generation=1585648129548914&amp;alt=media)\n\nWhy is that? Is it because fake faces much more diverse than real faces? How to deal with that?",
    "792969": "Did you undersample the fake samples?",
    "792972": "number of fake faces = number of real faces"
  },
  "source": "meta"
}