{
  "id": 98617,
  "title": "unfreeze() makes worse score?",
  "url": "/competitions/aptos2019-blindness-detection/discussion/98617",
  "author_name": "Takamichi Toda",
  "post_date": "2019-07-05T06:35:40.637000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I learned about image classification in the fast.ai course.</p>\n\n<p>The learned process was; First, train last layer by use freeze function. Second, checking loss and train all layers by use unfreeze function. In the second process, set the small learning rate for near input layer.</p>\n\n<p>This process was good to work in fast.ai's example task. However, It was not good in this competition.\nThe first process is no problem, but second process makes worse my score.</p>\n\n<p>My pre-train model is ResNet18.</p>\n\n<p>Are there same situation?</p>\n\n<p>Thanks</p>",
  "messages": [
    {
      "id": 568581,
      "postDate": "2019-07-05T06:35:40.637Z",
      "content": "<p>I learned about image classification in the fast.ai course.</p>\n\n<p>The learned process was; First, train last layer by use freeze function. Second, checking loss and train all layers by use unfreeze function. In the second process, set the small learning rate for near input layer.</p>\n\n<p>This process was good to work in fast.ai's example task. However, It was not good in this competition.\nThe first process is no problem, but second process makes worse my score.</p>\n\n<p>My pre-train model is ResNet18.</p>\n\n<p>Are there same situation?</p>\n\n<p>Thanks</p>",
      "rawMarkdown": "I learned about image classification in the fast.ai course.\n\nThe learned process was; First, train last layer by use freeze function. Second, checking loss and train all layers by use unfreeze function. In the second process, set the small learning rate for near input layer.\n\nThis process was good to work in fast.ai's example task. However, It was not good in this competition.\nThe first process is no problem, but second process makes worse my score.\n\nMy pre-train model is ResNet18.\n\nAre there same situation?\n\nThanks"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "568581": "I learned about image classification in the fast.ai course.\n\nThe learned process was; First, train last layer by use freeze function. Second, checking loss and train all layers by use unfreeze function. In the second process, set the small learning rate for near input layer.\n\nThis process was good to work in fast.ai's example task. However, It was not good in this competition.\nThe first process is no problem, but second process makes worse my score.\n\nMy pre-train model is ResNet18.\n\nAre there same situation?\n\nThanks"
  }
}