{
  "id": 130775,
  "title": "question about classification head in segmentation model",
  "url": "/competitions/understanding_cloud_organization/discussion/130775",
  "author_name": "",
  "post_date": "2020-02-16T08:41:01.128533700Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have benefited a lot from your excellent code. This is my first time to participate in the competition on image segmentation, and I see that the excellent code ranked at the top uses\nthe segmentation model with classification head. Can someone tell me How does the classification head improve the performance of the model in image segmentation and how should it be trained</p>",
  "messages": [
    {
      "id": "747315",
      "postDate": "02/16/2020 08:41:01",
      "content": "<p>I have benefited a lot from your excellent code. This is my first time to participate in the competition on image segmentation, and I see that the excellent code ranked at the top uses\nthe segmentation model with classification head. Can someone tell me How does the classification head improve the performance of the model in image segmentation and how should it be trained</p>",
      "rawMarkdown": "I have benefited a lot from your excellent code. This is my first time to participate in the competition on image segmentation, and I see that the excellent code ranked at the top uses\nthe segmentation model with classification head. Can someone tell me How does the classification head improve the performance of the model in image segmentation and how should it be trained",
      "votes": null
    },
    {
      "id": "748281",
      "postDate": "02/17/2020 10:49:19",
      "content": "<p>As you can see, the major task of this competition is to segment the cloud with certain shape. Under the major task, we can divide it into 2 sub task 1. classify the cloud through its shape 2. segment the cloud. So the cloud type classification is also important in this competition. But segmentation task is much complex than classification task, so we can add an addition classification head to help with pure classification job. So the weights in model will be updated by the gradient not only from segmentation head output but also classification head output.</p>",
      "rawMarkdown": "As you can see, the major task of this competition is to segment the cloud with certain shape. Under the major task, we can divide it into 2 sub task 1. classify the cloud through its shape 2. segment the cloud. So the cloud type classification is also important in this competition. But segmentation task is much complex than classification task, so we can add an addition classification head to help with pure classification job. So the weights in model will be updated by the gradient not only from segmentation head output but also classification head output.",
      "votes": null
    },
    {
      "id": "748429",
      "postDate": "02/17/2020 14:11:15",
      "content": "<p>I see some people training the classification head individually. And fix the weights of encoder and decoder. Is it right to make the classification head  better in this stage first, and then in the later overall training, the model will update encoder and decoder more according to the classification loss rather than the classification head? Do I understand that correctly?</p>",
      "rawMarkdown": "I see some people training the classification head individually. And fix the weights of encoder and decoder. Is it right to make the classification head  better in this stage first, and then in the later overall training, the model will update encoder and decoder more according to the classification loss rather than the classification head? Do I understand that correctly?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 748281,
      "author_name": "xiejialun",
      "author_url": "",
      "post_date": "02/17/2020 10:49:19",
      "content": "<p>As you can see, the major task of this competition is to segment the cloud with certain shape. Under the major task, we can divide it into 2 sub task 1. classify the cloud through its shape 2. segment the cloud. So the cloud type classification is also important in this competition. But segmentation task is much complex than classification task, so we can add an addition classification head to help with pure classification job. So the weights in model will be updated by the gradient not only from segmentation head output but also classification head output.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 748429,
      "author_name": "ssfailearning",
      "author_url": "",
      "post_date": "02/17/2020 14:11:15",
      "content": "<p>I see some people training the classification head individually. And fix the weights of encoder and decoder. Is it right to make the classification head  better in this stage first, and then in the later overall training, the model will update encoder and decoder more according to the classification loss rather than the classification head? Do I understand that correctly?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "747315": "I have benefited a lot from your excellent code. This is my first time to participate in the competition on image segmentation, and I see that the excellent code ranked at the top uses\nthe segmentation model with classification head. Can someone tell me How does the classification head improve the performance of the model in image segmentation and how should it be trained",
    "748281": "As you can see, the major task of this competition is to segment the cloud with certain shape. Under the major task, we can divide it into 2 sub task 1. classify the cloud through its shape 2. segment the cloud. So the cloud type classification is also important in this competition. But segmentation task is much complex than classification task, so we can add an addition classification head to help with pure classification job. So the weights in model will be updated by the gradient not only from segmentation head output but also classification head output.",
    "748429": "I see some people training the classification head individually. And fix the weights of encoder and decoder. Is it right to make the classification head  better in this stage first, and then in the later overall training, the model will update encoder and decoder more according to the classification loss rather than the classification head? Do I understand that correctly?"
  },
  "source": "meta"
}