{
  "id": 138772,
  "title": "What the best strategy to fine tune pretrained models?",
  "url": "/competitions/deepfake-detection-challenge/discussion/138772",
  "author_name": "",
  "post_date": "2020-03-26T08:48:58.384456400Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>What is the right approach that not cause overfitting? \nDo I need freezing some parts and train only final layers, then unfreeze and train whole model?\nOr train whole model right away?</p>",
  "messages": [
    {
      "id": "786835",
      "postDate": "03/26/2020 08:48:58",
      "content": "<p>What is the right approach that not cause overfitting? \nDo I need freezing some parts and train only final layers, then unfreeze and train whole model?\nOr train whole model right away?</p>",
      "rawMarkdown": "What is the right approach that not cause overfitting? \nDo I need freezing some parts and train only final layers, then unfreeze and train whole model?\nOr train whole model right away?",
      "votes": null
    },
    {
      "id": "786847",
      "postDate": "03/26/2020 09:05:10",
      "content": "<p>Train 2 epoches only(early stop) should give you a nice result.</p>",
      "rawMarkdown": "Train 2 epoches only(early stop) should give you a nice result.",
      "votes": null
    },
    {
      "id": "786886",
      "postDate": "03/26/2020 09:53:46",
      "content": "<p>I remember seeing on this forum people who trained for some significant number of epochs.\nLike 50 or so. How that even possible If networks overfitting is so fast?</p>",
      "rawMarkdown": "I remember seeing on this forum people who trained for some significant number of epochs.\nLike 50 or so. How that even possible If networks overfitting is so fast?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 786847,
      "author_name": "yuanzhezhou",
      "author_url": "",
      "post_date": "03/26/2020 09:05:10",
      "content": "<p>Train 2 epoches only(early stop) should give you a nice result.</p>",
      "votes": null,
      "replies": [
        {
          "id": 786886,
          "author_name": "mxmka87",
          "author_url": "",
          "post_date": "03/26/2020 09:53:46",
          "content": "<p>I remember seeing on this forum people who trained for some significant number of epochs.\nLike 50 or so. How that even possible If networks overfitting is so fast?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "786835": "What is the right approach that not cause overfitting? \nDo I need freezing some parts and train only final layers, then unfreeze and train whole model?\nOr train whole model right away?",
    "786847": "Train 2 epoches only(early stop) should give you a nice result.",
    "786886": "I remember seeing on this forum people who trained for some significant number of epochs.\nLike 50 or so. How that even possible If networks overfitting is so fast?"
  },
  "source": "meta"
}