{
  "id": 145768,
  "title": "why log loss, but no acc or other evaluation criteria",
  "url": "/competitions/deepfake-detection-challenge/discussion/145768",
  "author_name": "BokingChen",
  "post_date": "2020-04-24T13:10:16.005000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>The use of log loss to evaluate a model is a very unstable performance. Even if it is the same model, the local verification performance (acc) is basically the same, but the log loss of feedback is very different. Why not apply the relatively stable acc and other evaluation criteria</p>",
  "messages": [
    {
      "id": 819251,
      "postDate": "2020-04-24T13:10:16.007Z",
      "content": "<p>The use of log loss to evaluate a model is a very unstable performance. Even if it is the same model, the local verification performance (acc) is basically the same, but the log loss of feedback is very different. Why not apply the relatively stable acc and other evaluation criteria</p>",
      "rawMarkdown": "The use of log loss to evaluate a model is a very unstable performance. Even if it is the same model, the local verification performance (acc) is basically the same, but the log loss of feedback is very different. Why not apply the relatively stable acc and other evaluation criteria"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "819251": "The use of log loss to evaluate a model is a very unstable performance. Even if it is the same model, the local verification performance (acc) is basically the same, but the log loss of feedback is very different. Why not apply the relatively stable acc and other evaluation criteria"
  }
}