{
  "id": 142185,
  "title": "Suggestion about requirement of submitting pretrained models and codes for winners",
  "url": "/competitions/semi-inat-2020/discussion/142185",
  "author_name": "",
  "post_date": "2020-04-09T09:35:32.345124900Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Dear Organizer,</p>\n\n<p>Models pretrained on extra data (like NoisyStudent) can significantly boost the classification performance. It will be hard to guarantee the fairness without submitting pretrained models and codes for reimplementation after the competition. Is it possible to require all winners to submit pretrained models and codes please? In addition, it will be very helpful for other reseachers to further contribute to semi-supervised FGVC.</p>\n\n<p>Thank you very much!</p>",
  "messages": [
    {
      "id": "802257",
      "postDate": "04/09/2020 09:35:32",
      "content": "<p>Dear Organizer,</p>\n\n<p>Models pretrained on extra data (like NoisyStudent) can significantly boost the classification performance. It will be hard to guarantee the fairness without submitting pretrained models and codes for reimplementation after the competition. Is it possible to require all winners to submit pretrained models and codes please? In addition, it will be very helpful for other reseachers to further contribute to semi-supervised FGVC.</p>\n\n<p>Thank you very much!</p>",
      "rawMarkdown": "Dear Organizer,\n\nModels pretrained on extra data (like NoisyStudent) can significantly boost the classification performance. It will be hard to guarantee the fairness without submitting pretrained models and codes for reimplementation after the competition. Is it possible to require all winners to submit pretrained models and codes please? In addition, it will be very helpful for other reseachers to further contribute to semi-supervised FGVC.\n\nThank you very much!",
      "votes": null
    },
    {
      "id": "802701",
      "postDate": "04/09/2020 18:44:42",
      "content": "<p>Hi,\nThanks for the suggestion. Using extra data is prohibited except for ImageNet pre-trained models. However, since we are not using the Kaggle kernel we cannot prevent people from cheating. The winners will be invited to present their methods in the FGVC workshop, and we will think about asking the winner to provide their solutions.</p>",
      "rawMarkdown": "Hi,\nThanks for the suggestion. Using extra data is prohibited except for ImageNet pre-trained models. However, since we are not using the Kaggle kernel we cannot prevent people from cheating. The winners will be invited to present their methods in the FGVC workshop, and we will think about asking the winner to provide their solutions.",
      "votes": null
    },
    {
      "id": "802985",
      "postDate": "04/10/2020 02:57:34",
      "content": "<p>Hi Jong-Chyi,\nIt will be great to require winners to explicitly illustrate architectures and ImageNet-Top1-accuracies  of pretrained models they have used in their solutions. This requirement will be a step towards protecting the fairness. Many thanks! </p>",
      "rawMarkdown": "Hi Jong-Chyi,\nIt will be great to require winners to explicitly illustrate architectures and ImageNet-Top1-accuracies  of pretrained models they have used in their solutions. This requirement will be a step towards protecting the fairness. Many thanks!",
      "votes": null
    },
    {
      "id": "805020",
      "postDate": "04/12/2020 08:46:39",
      "content": "<p>Yes, that's a good point.</p>",
      "rawMarkdown": "Yes, that's a good point.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 802701,
      "author_name": "jcfredsu",
      "author_url": "",
      "post_date": "04/09/2020 18:44:42",
      "content": "<p>Hi,\nThanks for the suggestion. Using extra data is prohibited except for ImageNet pre-trained models. However, since we are not using the Kaggle kernel we cannot prevent people from cheating. The winners will be invited to present their methods in the FGVC workshop, and we will think about asking the winner to provide their solutions.</p>",
      "votes": null,
      "replies": [
        {
          "id": 802985,
          "author_name": "lyttonf",
          "author_url": "",
          "post_date": "04/10/2020 02:57:34",
          "content": "<p>Hi Jong-Chyi,\nIt will be great to require winners to explicitly illustrate architectures and ImageNet-Top1-accuracies  of pretrained models they have used in their solutions. This requirement will be a step towards protecting the fairness. Many thanks! </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 805020,
          "author_name": "sysuyanxp",
          "author_url": "",
          "post_date": "04/12/2020 08:46:39",
          "content": "<p>Yes, that's a good point.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "802257": "Dear Organizer,\n\nModels pretrained on extra data (like NoisyStudent) can significantly boost the classification performance. It will be hard to guarantee the fairness without submitting pretrained models and codes for reimplementation after the competition. Is it possible to require all winners to submit pretrained models and codes please? In addition, it will be very helpful for other reseachers to further contribute to semi-supervised FGVC.\n\nThank you very much!",
    "802701": "Hi,\nThanks for the suggestion. Using extra data is prohibited except for ImageNet pre-trained models. However, since we are not using the Kaggle kernel we cannot prevent people from cheating. The winners will be invited to present their methods in the FGVC workshop, and we will think about asking the winner to provide their solutions.",
    "802985": "Hi Jong-Chyi,\nIt will be great to require winners to explicitly illustrate architectures and ImageNet-Top1-accuracies  of pretrained models they have used in their solutions. This requirement will be a step towards protecting the fairness. Many thanks!",
    "805020": "Yes, that's a good point."
  },
  "source": "meta"
}