{
  "id": 70585,
  "title": "post processing",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/70585",
  "author_name": "hengck23",
  "post_date": "2018-11-05T12:47:01.167000",
  "votes": 12,
  "comment_count": 9,
  "views": 0,
  "content": "<p>this is maybe the key to wining, see attached image:</p>\n\n<p>Also cross entropy is not the best loss. We can make three guesses. Your loss should take advantage of that</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/415648/10611/post-processing.png\" alt=\"enter image description here\"></p>",
  "messages": [
    {
      "id": 415648,
      "postDate": "2018-11-05T12:47:01.167Z",
      "content": "<p>this is maybe the key to wining, see attached image:</p>\n\n<p>Also cross entropy is not the best loss. We can make three guesses. Your loss should take advantage of that</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/415648/10611/post-processing.png\" alt=\"enter image description here\"></p>",
      "rawMarkdown": "this is maybe the key to wining, see attached image:\n\nAlso cross entropy is not the best loss. We can make three guesses. Your loss should take advantage of that\n\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/415648/10611/post-processing.png",
      "votes": 12
    },
    {
      "id": 415650,
      "postDate": "2018-11-05T12:54:49.120Z",
      "content": "<p>e.g softmax output = (0,1). prob values is limited to 0 to one and sum(Pi)=1</p>\n\n<p>what happens if you modify it to  Pi between (0, 0.75) and sum(Pi)=1. The CNN is forced to learn a second best guess.</p>",
      "rawMarkdown": "e.g softmax output = (0,1). prob values is limited to 0 to one and sum(Pi)=1\n\nwhat happens if you modify it to  Pi between (0, 0.75) and sum(Pi)=1. The CNN is forced to learn a second best guess.\n",
      "votes": 2,
      "replies": [
        {
          "id": 415974,
          "postDate": "2018-11-06T01:46:14.847Z",
          "content": "<p>Hi, How can we set the second best guess supervision signal? since the ground truth label is only one class. </p>",
          "rawMarkdown": "Hi, How can we set the second best guess supervision signal? since the ground truth label is only one class. "
        },
        {
          "id": 416320,
          "postDate": "2018-11-06T14:39:45.667Z",
          "content": "<p>maybe some ranking network that predict:</p>\n\n<p>(image,A,B,C) &gt; (image,A,C,B)?</p>",
          "rawMarkdown": "maybe some ranking network that predict:\n\n(image,A,B,C) &gt; (image,A,C,B)?"
        },
        {
          "id": 417251,
          "postDate": "2018-11-08T01:42:29.410Z",
          "content": "<p>assuming you network output  3 guesses (e.g. using lstm seq-to-seq), p_1, p_2, p_3</p>\n\n<p>each p_i is of dim = 340 = num of classes</p>\n\n<p>the loss can be:</p>\n\n<p>-log(p_1,true) -  (1-p_1,true)*log(p_2,true) -  (1-p_1,true)*(1-p_2,true)*log(p_3,true)</p>",
          "rawMarkdown": "assuming you network output  3 guesses (e.g. using lstm seq-to-seq), p_1, p_2, p_3\n\neach p_i is of dim = 340 = num of classes\n\nthe loss can be:\n\n-log(p_1,true) -  (1-p_1,true)*log(p_2,true) -  (1-p_1,true)*(1-p_2,true)*log(p_3,true)\n\n",
          "votes": 1
        }
      ]
    },
    {
      "id": 416323,
      "postDate": "2018-11-06T14:44:42.410Z",
      "content": "<p>Improving Pairwise Ranking for Multi-label Image Classification\n<a href=\"https://arxiv.org/pdf/1704.03135.pdf\">https://arxiv.org/pdf/1704.03135.pdf</a></p>",
      "rawMarkdown": "Improving Pairwise Ranking for Multi-label Image Classification\nhttps://arxiv.org/pdf/1704.03135.pdf"
    },
    {
      "id": 415794,
      "postDate": "2018-11-05T17:55:22.277Z",
      "content": "<p>References </p>\n\n<p><a href=\"https://github.com/arijitx/Amazon-Satelite-Image-Labeling\">https://github.com/arijitx/Amazon-Satelite-Image-Labeling</a></p>",
      "rawMarkdown": "References \n\nhttps://github.com/arijitx/Amazon-Satelite-Image-Labeling"
    },
    {
      "id": 415696,
      "postDate": "2018-11-05T14:12:53.820Z",
      "content": "<p>If some test images are repeated, the pred from same model should be same. right?</p>",
      "rawMarkdown": "If some test images are repeated, the pred from same model should be same. right?",
      "replies": [
        {
          "id": 415700,
          "postDate": "2018-11-05T14:15:59.097Z",
          "content": "<p>no necessary.</p>\n\n<p>if your prediction is A(0.2),B(0.2),C,</p>\n\n<p>you can modify and put A,B,C in one image and B,A,C in another</p>",
          "rawMarkdown": "no necessary.\n\nif your prediction is A(0.2),B(0.2),C,\n\nyou can modify and put A,B,C in one image and B,A,C in another"
        },
        {
          "id": 416306,
          "postDate": "2018-11-06T14:23:46.253Z",
          "content": "<p>I got it. If A and B have close scores, we can switch the positions of A/B, and the positions of C don't have to be modified. right?</p>",
          "rawMarkdown": "I got it. If A and B have close scores, we can switch the positions of A/B, and the positions of C don't have to be modified. right?"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 415650,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-11-05T12:54:49.120000",
      "content": "<p>e.g softmax output = (0,1). prob values is limited to 0 to one and sum(Pi)=1</p>\n\n<p>what happens if you modify it to  Pi between (0, 0.75) and sum(Pi)=1. The CNN is forced to learn a second best guess.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 415974,
          "author_name": "good good study",
          "author_url": "",
          "post_date": "2018-11-06T01:46:14.847000",
          "content": "<p>Hi, How can we set the second best guess supervision signal? since the ground truth label is only one class. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 416320,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2018-11-06T14:39:45.667000",
          "content": "<p>maybe some ranking network that predict:</p>\n\n<p>(image,A,B,C) &gt; (image,A,C,B)?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 417251,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2018-11-08T01:42:29.410000",
          "content": "<p>assuming you network output  3 guesses (e.g. using lstm seq-to-seq), p_1, p_2, p_3</p>\n\n<p>each p_i is of dim = 340 = num of classes</p>\n\n<p>the loss can be:</p>\n\n<p>-log(p_1,true) -  (1-p_1,true)*log(p_2,true) -  (1-p_1,true)*(1-p_2,true)*log(p_3,true)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 416323,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-11-06T14:44:42.410000",
      "content": "<p>Improving Pairwise Ranking for Multi-label Image Classification\n<a href=\"https://arxiv.org/pdf/1704.03135.pdf\">https://arxiv.org/pdf/1704.03135.pdf</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 415794,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-11-05T17:55:22.277000",
      "content": "<p>References </p>\n\n<p><a href=\"https://github.com/arijitx/Amazon-Satelite-Image-Labeling\">https://github.com/arijitx/Amazon-Satelite-Image-Labeling</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 415696,
      "author_name": "Gary",
      "author_url": "",
      "post_date": "2018-11-05T14:12:53.820000",
      "content": "<p>If some test images are repeated, the pred from same model should be same. right?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 415700,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2018-11-05T14:15:59.097000",
          "content": "<p>no necessary.</p>\n\n<p>if your prediction is A(0.2),B(0.2),C,</p>\n\n<p>you can modify and put A,B,C in one image and B,A,C in another</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 416306,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2018-11-06T14:23:46.253000",
          "content": "<p>I got it. If A and B have close scores, we can switch the positions of A/B, and the positions of C don't have to be modified. right?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "415648": "this is maybe the key to wining, see attached image:\n\nAlso cross entropy is not the best loss. We can make three guesses. Your loss should take advantage of that\n\n\n  ![enter image description here][1]\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/415648/10611/post-processing.png",
    "415650": "e.g softmax output = (0,1). prob values is limited to 0 to one and sum(Pi)=1\n\nwhat happens if you modify it to  Pi between (0, 0.75) and sum(Pi)=1. The CNN is forced to learn a second best guess.\n",
    "416323": "Improving Pairwise Ranking for Multi-label Image Classification\nhttps://arxiv.org/pdf/1704.03135.pdf",
    "415794": "References \n\nhttps://github.com/arijitx/Amazon-Satelite-Image-Labeling",
    "415696": "If some test images are repeated, the pred from same model should be same. right?"
  }
}