{
  "id": 138187,
  "title": "small and clean train set or big but noisy?",
  "url": "/competitions/deepfake-detection-challenge/discussion/138187",
  "author_name": "ZhiHao Wang",
  "post_date": "2020-03-24T03:47:54.165000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>There are about 12w videos, and every video has about 300 frames, so the total data set is 36 milliion. but I only have a RTX2080TI, so I take 10 frames every video and use all videos reached about 1.2 million to train. But my score stoped at 0.38. \nAnd I found that many people used just 40000 videos(real:fake=1:1). 10 frames for one video.\nSo I want to know what's the best split if don't have multi-GPUS.</p>",
  "messages": [
    {
      "id": 784259,
      "postDate": "2020-03-24T03:47:54.167Z",
      "content": "<p>There are about 12w videos, and every video has about 300 frames, so the total data set is 36 milliion. but I only have a RTX2080TI, so I take 10 frames every video and use all videos reached about 1.2 million to train. But my score stoped at 0.38. \nAnd I found that many people used just 40000 videos(real:fake=1:1). 10 frames for one video.\nSo I want to know what's the best split if don't have multi-GPUS.</p>",
      "rawMarkdown": "There are about 12w videos, and every video has about 300 frames, so the total data set is 36 milliion. but I only have a RTX2080TI, so I take 10 frames every video and use all videos reached about 1.2 million to train. But my score stoped at 0.38. \nAnd I found that many people used just 40000 videos(real:fake=1:1). 10 frames for one video.\nSo I want to know what's the best split if don't have multi-GPUS.",
      "votes": 2
    },
    {
      "id": 784288,
      "postDate": "2020-03-24T04:47:28.127Z",
      "content": "<p>I have a devbox with 1 GTX 1080 Ti which I'm using for training and 1 GTX 1080 which I'm using for some experiments. Think that such setup is pretty enough to get 0.2-0.3 public lb score. Keep trying )</p>",
      "rawMarkdown": "I have a devbox with 1 GTX 1080 Ti which I'm using for training and 1 GTX 1080 which I'm using for some experiments. Think that such setup is pretty enough to get 0.2-0.3 public lb score. Keep trying )",
      "votes": 5,
      "isDeleted": true,
      "replies": [
        {
          "id": 784292,
          "postDate": "2020-03-24T04:52:37.003Z",
          "content": "<p>Thanks</p>",
          "rawMarkdown": "Thanks"
        },
        {
          "id": 784383,
          "postDate": "2020-03-24T07:07:55.267Z",
          "content": "<p><a href=\"/vladislavleketush\">@vladislavleketush</a> Do you clean the dataset?</p>",
          "rawMarkdown": "@vladislavleketush Do you clean the dataset?\n"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 784288,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-03-24T04:47:28.127000",
      "content": "<p>I have a devbox with 1 GTX 1080 Ti which I'm using for training and 1 GTX 1080 which I'm using for some experiments. Think that such setup is pretty enough to get 0.2-0.3 public lb score. Keep trying )</p>",
      "votes": 5,
      "replies": [
        {
          "id": 784292,
          "author_name": "ZhiHao Wang",
          "author_url": "",
          "post_date": "2020-03-24T04:52:37.003000",
          "content": "<p>Thanks</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 784383,
          "author_name": "BokingChen",
          "author_url": "",
          "post_date": "2020-03-24T07:07:55.267000",
          "content": "<p><a href=\"/vladislavleketush\">@vladislavleketush</a> Do you clean the dataset?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "784259": "There are about 12w videos, and every video has about 300 frames, so the total data set is 36 milliion. but I only have a RTX2080TI, so I take 10 frames every video and use all videos reached about 1.2 million to train. But my score stoped at 0.38. \nAnd I found that many people used just 40000 videos(real:fake=1:1). 10 frames for one video.\nSo I want to know what's the best split if don't have multi-GPUS.",
    "784288": "I have a devbox with 1 GTX 1080 Ti which I'm using for training and 1 GTX 1080 which I'm using for some experiments. Think that such setup is pretty enough to get 0.2-0.3 public lb score. Keep trying )"
  }
}