{
  "id": 130124,
  "title": "Ratio of True to Fake in Test set",
  "url": "/competitions/deepfake-detection-challenge/discussion/130124",
  "author_name": "",
  "post_date": "2020-02-12T08:57:51.087960300Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi! New to this competition. I looked around the posts and saw that during training, some high-scoring solutions undersampled fake frames to speed up training. Wouldn't this be harmful if ratio of true:fake in private LB is not 1:1? </p>\n\n<p>Also, I wonder what is the ratio of true:fake for public / private LB??\nDid you undersample fake frames?\nAny help appreciated, thanks</p>",
  "messages": [
    {
      "id": "743743",
      "postDate": "02/12/2020 08:57:51",
      "content": "<p>Hi! New to this competition. I looked around the posts and saw that during training, some high-scoring solutions undersampled fake frames to speed up training. Wouldn't this be harmful if ratio of true:fake in private LB is not 1:1? </p>\n\n<p>Also, I wonder what is the ratio of true:fake for public / private LB??\nDid you undersample fake frames?\nAny help appreciated, thanks</p>",
      "rawMarkdown": "Hi! New to this competition. I looked around the posts and saw that during training, some high-scoring solutions undersampled fake frames to speed up training. Wouldn't this be harmful if ratio of true:fake in private LB is not 1:1? \n\nAlso, I wonder what is the ratio of true:fake for public / private LB??\nDid you undersample fake frames?\nAny help appreciated, thanks",
      "votes": null
    },
    {
      "id": "743772",
      "postDate": "02/12/2020 09:43:13",
      "content": "<p>I  also  want  to  know  about  it .</p>",
      "rawMarkdown": "I  also  want  to  know  about  it .",
      "votes": null
    },
    {
      "id": "743906",
      "postDate": "02/12/2020 11:33:31",
      "content": "<p>You can have a look here: <a href=\"https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121348\">https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121348</a> \nPrivate LB seems to have 1:1 ratio between true:fake using information from log loss. \nEven if there wasn't an equal ratio, it's still better to train with a balanced dataset.</p>",
      "rawMarkdown": "You can have a look here: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121348 \nPrivate LB seems to have 1:1 ratio between true:fake using information from log loss. \nEven if there wasn't an equal ratio, it's still better to train with a balanced dataset.",
      "votes": null
    },
    {
      "id": "744528",
      "postDate": "02/12/2020 22:54:22",
      "content": "<p>It is 1:1 in the public LB.\n<a href=\"https://www.kaggle.com/c/deepfake-detection-challenge/discussion/126524\">https://www.kaggle.com/c/deepfake-detection-challenge/discussion/126524</a></p>",
      "rawMarkdown": "It is 1:1 in the public LB.\nhttps://www.kaggle.com/c/deepfake-detection-challenge/discussion/126524",
      "votes": null
    },
    {
      "id": "754727",
      "postDate": "02/24/2020 01:24:07",
      "content": "<p><a href=\"/petewills\">@petewills</a> <a href=\"/rafiko1\">@rafiko1</a> Thank you guys very much! Makes things much clearer for me :)</p>",
      "rawMarkdown": "petewills @rafiko1 Thank you guys very much! Makes things much clearer for me :)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 743772,
      "author_name": "xujingzhao",
      "author_url": "",
      "post_date": "02/12/2020 09:43:13",
      "content": "<p>I  also  want  to  know  about  it .</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 743906,
      "author_name": "rafiko1",
      "author_url": "",
      "post_date": "02/12/2020 11:33:31",
      "content": "<p>You can have a look here: <a href=\"https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121348\">https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121348</a> \nPrivate LB seems to have 1:1 ratio between true:fake using information from log loss. \nEven if there wasn't an equal ratio, it's still better to train with a balanced dataset.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 744528,
      "author_name": "petewills",
      "author_url": "",
      "post_date": "02/12/2020 22:54:22",
      "content": "<p>It is 1:1 in the public LB.\n<a href=\"https://www.kaggle.com/c/deepfake-detection-challenge/discussion/126524\">https://www.kaggle.com/c/deepfake-detection-challenge/discussion/126524</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 754727,
      "author_name": "roguekk007",
      "author_url": "",
      "post_date": "02/24/2020 01:24:07",
      "content": "<p><a href=\"/petewills\">@petewills</a> <a href=\"/rafiko1\">@rafiko1</a> Thank you guys very much! Makes things much clearer for me :)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "743743": "Hi! New to this competition. I looked around the posts and saw that during training, some high-scoring solutions undersampled fake frames to speed up training. Wouldn't this be harmful if ratio of true:fake in private LB is not 1:1? \n\nAlso, I wonder what is the ratio of true:fake for public / private LB??\nDid you undersample fake frames?\nAny help appreciated, thanks",
    "743772": "I  also  want  to  know  about  it .",
    "743906": "You can have a look here: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121348 \nPrivate LB seems to have 1:1 ratio between true:fake using information from log loss. \nEven if there wasn't an equal ratio, it's still better to train with a balanced dataset.",
    "744528": "It is 1:1 in the public LB.\nhttps://www.kaggle.com/c/deepfake-detection-challenge/discussion/126524",
    "754727": "petewills @rafiko1 Thank you guys very much! Makes things much clearer for me :)"
  },
  "source": "meta"
}