{
  "id": 105873,
  "title": "\"Freezing\" layers in a model",
  "url": "/competitions/aptos2019-blindness-detection/discussion/105873",
  "author_name": "Benson Jin",
  "post_date": "2019-08-27T00:04:15.148000",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi I am currently ensembling different models for this competition. However, I was looking through some kernels and people were \"freezing\" layers and then \"unfreezing\" them.  What exactly is the point of this? Sorry, if I am being annoying. </p>",
  "messages": [
    {
      "id": 608557,
      "postDate": "2019-08-27T00:04:15.147Z",
      "content": "<p>Hi I am currently ensembling different models for this competition. However, I was looking through some kernels and people were \"freezing\" layers and then \"unfreezing\" them.  What exactly is the point of this? Sorry, if I am being annoying. </p>",
      "rawMarkdown": "Hi I am currently ensembling different models for this competition. However, I was looking through some kernels and people were \"freezing\" layers and then \"unfreezing\" them.  What exactly is the point of this? Sorry, if I am being annoying. ",
      "votes": 1
    },
    {
      "id": 610417,
      "postDate": "2019-08-28T19:26:50.860Z",
      "content": "<p>If you have a pre-trained model, and you want to adapt it on your task, you usually freeze the first layers because they capture general features, you train just the last layers because they are specific to each task. you can after that fine-tune your model by unfreezing all layers and do training with a very small learning-rate.</p>",
      "rawMarkdown": "If you have a pre-trained model, and you want to adapt it on your task, you usually freeze the first layers because they capture general features, you train just the last layers because they are specific to each task. you can after that fine-tune your model by unfreezing all layers and do training with a very small learning-rate."
    },
    {
      "id": 608565,
      "postDate": "2019-08-27T00:26:02.863Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true,
      "replies": [
        {
          "id": 608622,
          "postDate": "2019-08-27T03:07:56.347Z",
          "content": "<p>Thank you so much!   :) </p>",
          "rawMarkdown": "Thank you so much!   :) "
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 610417,
      "author_name": "JM100",
      "author_url": "",
      "post_date": "2019-08-28T19:26:50.860000",
      "content": "<p>If you have a pre-trained model, and you want to adapt it on your task, you usually freeze the first layers because they capture general features, you train just the last layers because they are specific to each task. you can after that fine-tune your model by unfreezing all layers and do training with a very small learning-rate.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 608565,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-08-27T00:26:02.863000",
      "content": "",
      "votes": 1,
      "replies": [
        {
          "id": 608622,
          "author_name": "Benson Jin",
          "author_url": "",
          "post_date": "2019-08-27T03:07:56.347000",
          "content": "<p>Thank you so much!   :) </p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "608557": "Hi I am currently ensembling different models for this competition. However, I was looking through some kernels and people were \"freezing\" layers and then \"unfreezing\" them.  What exactly is the point of this? Sorry, if I am being annoying. ",
    "610417": "If you have a pre-trained model, and you want to adapt it on your task, you usually freeze the first layers because they capture general features, you train just the last layers because they are specific to each task. you can after that fine-tune your model by unfreezing all layers and do training with a very small learning-rate.",
    "608565": ""
  }
}