{
  "id": 136886,
  "title": "Conclusion from Top Solutions: Solutiosn for Unseen Graphemes",
  "url": "/competitions/bengaliai-cv19/discussion/136886",
  "author_name": "",
  "post_date": "2020-03-18T06:24:28.123021300Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>It seems that unseen graphemes are the biggest issue in the shake. I observed a few major trends in top solutions that dealt with this problem:\n- simple models (densenet, efficientnet, etc.)\n- CAMCutmix/ZoneCutmix\n- class balance based on label frequency</p>\n\n<p>If you observed anything else, please comment below.</p>\n\n<p>Thanks!</p>",
  "messages": [
    {
      "id": "778092",
      "postDate": "03/18/2020 06:24:28",
      "content": "<p>It seems that unseen graphemes are the biggest issue in the shake. I observed a few major trends in top solutions that dealt with this problem:\n- simple models (densenet, efficientnet, etc.)\n- CAMCutmix/ZoneCutmix\n- class balance based on label frequency</p>\n\n<p>If you observed anything else, please comment below.</p>\n\n<p>Thanks!</p>",
      "rawMarkdown": "It seems that unseen graphemes are the biggest issue in the shake. I observed a few major trends in top solutions that dealt with this problem:\n- simple models (densenet, efficientnet, etc.)\n- CAMCutmix/ZoneCutmix\n- class balance based on label frequency\n\nIf you observed anything else, please comment below.\n\nThanks!",
      "votes": null
    },
    {
      "id": "778114",
      "postDate": "03/18/2020 06:53:48",
      "content": "<p>Found a similar (better) thread here: <a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/136030\">https://www.kaggle.com/c/bengaliai-cv19/discussion/136030</a></p>",
      "rawMarkdown": "Found a similar (better) thread here: https://www.kaggle.com/c/bengaliai-cv19/discussion/136030",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 778114,
      "author_name": "tonychenxyz",
      "author_url": "",
      "post_date": "03/18/2020 06:53:48",
      "content": "<p>Found a similar (better) thread here: <a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/136030\">https://www.kaggle.com/c/bengaliai-cv19/discussion/136030</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "778092": "It seems that unseen graphemes are the biggest issue in the shake. I observed a few major trends in top solutions that dealt with this problem:\n- simple models (densenet, efficientnet, etc.)\n- CAMCutmix/ZoneCutmix\n- class balance based on label frequency\n\nIf you observed anything else, please comment below.\n\nThanks!",
    "778114": "Found a similar (better) thread here: https://www.kaggle.com/c/bengaliai-cv19/discussion/136030"
  },
  "source": "meta"
}