{
  "id": 72095,
  "title": "Micro f1 score troubled me in training.",
  "url": "/competitions/human-protein-atlas-image-classification/discussion/72095",
  "author_name": "",
  "post_date": "2018-11-20T08:48:41.307000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I used resnet50 as the basic model but change the last fully connection to 28 dims. Starting training with adam ,lr=1e-4. but I found that in many classes the f1 score equal to .0, I puzzled that if it's a common phenomenon in the begining of training, or if there are some problems in my process. help,</p>",
  "messages": [
    {
      "id": 424513,
      "postDate": "2018-11-20T08:48:41.307Z",
      "content": "<p>I used resnet50 as the basic model but change the last fully connection to 28 dims. Starting training with adam ,lr=1e-4. but I found that in many classes the f1 score equal to .0, I puzzled that if it's a common phenomenon in the begining of training, or if there are some problems in my process. help,</p>",
      "rawMarkdown": "I used resnet50 as the basic model but change the last fully connection to 28 dims. Starting training with adam ,lr=1e-4. but I found that in many classes the f1 score equal to .0, I puzzled that if it's a common phenomenon in the begining of training, or if there are some problems in my process. help,",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "424513": "I used resnet50 as the basic model but change the last fully connection to 28 dims. Starting training with adam ,lr=1e-4. but I found that in many classes the f1 score equal to .0, I puzzled that if it's a common phenomenon in the begining of training, or if there are some problems in my process. help,"
  }
}