{
  "id": 39776,
  "title": "auxillary loss for hierarchies + evaluation",
  "url": "/competitions/cdiscount-image-classification-challenge/discussion/39776",
  "author_name": "",
  "post_date": "2017-09-20T21:09:36.051897Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>hey, \nhas anybody thought about using level 2 and level 3 descriptions to enhance training? i thought of adding an auxillary loss to a lower mid-level layer @vgg. i guess this should make 5000-class optimization more feasible and get you a better gradient and more generalization. </p>\n\n<p>additionally one q: is the evaluation just accuracy or per class accuracy? i'm reading it as just accuracy?</p>",
  "messages": [
    {
      "id": "223009",
      "postDate": "09/20/2017 21:09:36",
      "content": "<p>hey, \nhas anybody thought about using level 2 and level 3 descriptions to enhance training? i thought of adding an auxillary loss to a lower mid-level layer @vgg. i guess this should make 5000-class optimization more feasible and get you a better gradient and more generalization. </p>\n\n<p>additionally one q: is the evaluation just accuracy or per class accuracy? i'm reading it as just accuracy?</p>",
      "rawMarkdown": "hey, \nhas anybody thought about using level 2 and level 3 descriptions to enhance training? i thought of adding an auxillary loss to a lower mid-level layer @vgg. i guess this should make 5000-class optimization more feasible and get you a better gradient and more generalization. \n\nadditionally one q: is the evaluation just accuracy or per class accuracy? i'm reading it as just accuracy?",
      "votes": null
    },
    {
      "id": "223013",
      "postDate": "09/20/2017 21:22:33",
      "content": "<p>It's overall accuracy. In other words, <code>rows_correctly_classified / total_rows</code>.</p>",
      "rawMarkdown": "It's overall accuracy. In other words, `rows_correctly_classified / total_rows`.",
      "votes": null
    },
    {
      "id": "231539",
      "postDate": "10/15/2017 09:29:52",
      "content": "<p>I was thinking about something. If we are able to correctly draw the hiearchical tree of classes. Could we train a network not to predict a vector valued of probability but a probability tree under constrained. I think it could help with this kind of large class pb .</p>",
      "rawMarkdown": "I was thinking about something. If we are able to correctly draw the hiearchical tree of classes. Could we train a network not to predict a vector valued of probability but a probability tree under constrained. I think it could help with this kind of large class pb .",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 223013,
      "author_name": "inversion",
      "author_url": "",
      "post_date": "09/20/2017 21:22:33",
      "content": "<p>It's overall accuracy. In other words, <code>rows_correctly_classified / total_rows</code>.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 231539,
      "author_name": "tchaton",
      "author_url": "",
      "post_date": "10/15/2017 09:29:52",
      "content": "<p>I was thinking about something. If we are able to correctly draw the hiearchical tree of classes. Could we train a network not to predict a vector valued of probability but a probability tree under constrained. I think it could help with this kind of large class pb .</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "223009": "hey, \nhas anybody thought about using level 2 and level 3 descriptions to enhance training? i thought of adding an auxillary loss to a lower mid-level layer @vgg. i guess this should make 5000-class optimization more feasible and get you a better gradient and more generalization. \n\nadditionally one q: is the evaluation just accuracy or per class accuracy? i'm reading it as just accuracy?",
    "223013": "It's overall accuracy. In other words, `rows_correctly_classified / total_rows`.",
    "231539": "I was thinking about something. If we are able to correctly draw the hiearchical tree of classes. Could we train a network not to predict a vector valued of probability but a probability tree under constrained. I think it could help with this kind of large class pb ."
  },
  "source": "meta"
}