{
  "id": 326285,
  "title": "step by step defreezing layers (idea)",
  "url": "/competitions/herbarium-2022-fgvc9/discussion/326285",
  "author_name": "Eugene",
  "post_date": "2022-05-21T08:55:21.901000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello guys.<br>\nI'm trying to train mobilenet v2. <br>\nFirst idea was to freeze all layers except classifier. But the model didn't seem to learn anything.<br>\nThen I added one extra layer in classifier, so now the classifier looks like:<br>\nlinear(1028 * 7000), linear(7000 * 15505)<br>\nBut loss didn't fall less than 7.0<br>\nSo, I've decided to defreeze pref 2 conv layers and now model is able to decrees loss up to 6.0 but can't learn more.<br>\nWhat do u thing is it rational to defreeze layer by layer? </p>",
  "messages": [
    {
      "id": 1796870,
      "postDate": "2022-05-21T08:55:21.903Z",
      "content": "<p>Hello guys.<br>\nI'm trying to train mobilenet v2. <br>\nFirst idea was to freeze all layers except classifier. But the model didn't seem to learn anything.<br>\nThen I added one extra layer in classifier, so now the classifier looks like:<br>\nlinear(1028 * 7000), linear(7000 * 15505)<br>\nBut loss didn't fall less than 7.0<br>\nSo, I've decided to defreeze pref 2 conv layers and now model is able to decrees loss up to 6.0 but can't learn more.<br>\nWhat do u thing is it rational to defreeze layer by layer? </p>",
      "rawMarkdown": "Hello guys.\nI'm trying to train mobilenet v2. \nFirst idea was to freeze all layers except classifier. But the model didn't seem to learn anything.\nThen I added one extra layer in classifier, so now the classifier looks like:\nlinear(1028 * 7000), linear(7000 * 15505)\nBut loss didn't fall less than 7.0\nSo, I've decided to defreeze pref 2 conv layers and now model is able to decrees loss up to 6.0 but can't learn more.\nWhat do u thing is it rational to defreeze layer by layer? ",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1796870": "Hello guys.\nI'm trying to train mobilenet v2. \nFirst idea was to freeze all layers except classifier. But the model didn't seem to learn anything.\nThen I added one extra layer in classifier, so now the classifier looks like:\nlinear(1028 * 7000), linear(7000 * 15505)\nBut loss didn't fall less than 7.0\nSo, I've decided to defreeze pref 2 conv layers and now model is able to decrees loss up to 6.0 but can't learn more.\nWhat do u thing is it rational to defreeze layer by layer? "
  }
}