{
  "id": 44132,
  "title": "Has anyone trained hierarchy classification net in yolo9000 successfully?",
  "url": "/competitions/cdiscount-image-classification-challenge/discussion/44132",
  "author_name": "",
  "post_date": "2017-11-24T02:24:07.272148Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have been trying applying hierarchy classification to resnet recently, but it sames that there's always some problems. </p>\n\n<p>I splited total classes into 49 top-level classes, so there are 49+1 softmax. I added cross-entropy for each softmax to make up the total loss. But the net didn't tend to converge after some epoches.</p>\n\n<p>Is there something wrong in my implemention? Thanks in advance.</p>",
  "messages": [
    {
      "id": "247829",
      "postDate": "11/24/2017 02:24:07",
      "content": "<p>I have been trying applying hierarchy classification to resnet recently, but it sames that there's always some problems. </p>\n\n<p>I splited total classes into 49 top-level classes, so there are 49+1 softmax. I added cross-entropy for each softmax to make up the total loss. But the net didn't tend to converge after some epoches.</p>\n\n<p>Is there something wrong in my implemention? Thanks in advance.</p>",
      "rawMarkdown": "I have been trying applying hierarchy classification to resnet recently, but it sames that there's always some problems. \n\nI splited total classes into 49 top-level classes, so there are 49+1 softmax. I added cross-entropy for each softmax to make up the total loss. But the net didn't tend to converge after some epoches.\n\nIs there something wrong in my implemention? Thanks in advance.",
      "votes": null
    },
    {
      "id": "247874",
      "postDate": "11/24/2017 07:01:27",
      "content": "<p>I think hierarchy loose accuracy. Its good concept for yolo9000. If they open prepared dataset we can train better accuracy without hierarchy =)</p>",
      "rawMarkdown": "I think hierarchy loose accuracy. Its good concept for yolo9000. If they open prepared dataset we can train better accuracy without hierarchy =)",
      "votes": null
    },
    {
      "id": "247882",
      "postDate": "11/24/2017 08:09:41",
      "content": "<p>so I guess you have came up with some other ideas to deal with these huge number of categories?</p>",
      "rawMarkdown": "so I guess you have came up with some other ideas to deal with these huge number of categories?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 247874,
      "author_name": "nicksergievskiy",
      "author_url": "",
      "post_date": "11/24/2017 07:01:27",
      "content": "<p>I think hierarchy loose accuracy. Its good concept for yolo9000. If they open prepared dataset we can train better accuracy without hierarchy =)</p>",
      "votes": null,
      "replies": [
        {
          "id": 247882,
          "author_name": "plainsailing",
          "author_url": "",
          "post_date": "11/24/2017 08:09:41",
          "content": "<p>so I guess you have came up with some other ideas to deal with these huge number of categories?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "247829": "I have been trying applying hierarchy classification to resnet recently, but it sames that there's always some problems. \n\nI splited total classes into 49 top-level classes, so there are 49+1 softmax. I added cross-entropy for each softmax to make up the total loss. But the net didn't tend to converge after some epoches.\n\nIs there something wrong in my implemention? Thanks in advance.",
    "247874": "I think hierarchy loose accuracy. Its good concept for yolo9000. If they open prepared dataset we can train better accuracy without hierarchy =)",
    "247882": "so I guess you have came up with some other ideas to deal with these huge number of categories?"
  },
  "source": "meta"
}