{
  "id": 108227,
  "title": "[UPDATED] Tricks I wish I have tried",
  "url": "/competitions/aptos2019-blindness-detection/discussion/108227",
  "author_name": "",
  "post_date": "2019-09-10T06:32:42.385365800Z",
  "votes": 6,
  "comment_count": 8,
  "views": 0,
  "content": "<p>Let's share the ideas you came up with but you haven't had time or skill to try or they simply didn't work for some reason. </p>\n\n<p>To keep this topic rather dense lets avoid what's already mentioned in other threads (ensemble, pseudolabels, etc.) </p>\n\n<p>For instance optimizing <a href=\"https://www.kaggle.com/zaharch/minimizing-qwk-directly-with-nn\">kappa loss</a> directly didn't work for me (2x runs with 2 folds each)</p>\n\n<p>[UPDATE] here I provide the list of tricks mentioned in this thread\n- Optimize kappa loss directly\n- Attention Network - <a href=\"https://eccv2018.org/openaccess/content_ECCV_2018/papers/Sanghyun_Woo_Convolutional_Block_Attention_ECCV_2018_paper.pdf\">CBAM</a></p>",
  "messages": [
    {
      "id": "622844",
      "postDate": "09/10/2019 06:32:42",
      "content": "<p>Let's share the ideas you came up with but you haven't had time or skill to try or they simply didn't work for some reason. </p>\n\n<p>To keep this topic rather dense lets avoid what's already mentioned in other threads (ensemble, pseudolabels, etc.) </p>\n\n<p>For instance optimizing <a href=\"https://www.kaggle.com/zaharch/minimizing-qwk-directly-with-nn\">kappa loss</a> directly didn't work for me (2x runs with 2 folds each)</p>\n\n<p>[UPDATE] here I provide the list of tricks mentioned in this thread\n- Optimize kappa loss directly\n- Attention Network - <a href=\"https://eccv2018.org/openaccess/content_ECCV_2018/papers/Sanghyun_Woo_Convolutional_Block_Attention_ECCV_2018_paper.pdf\">CBAM</a></p>",
      "rawMarkdown": "Let's share the ideas you came up with but you haven't had time or skill to try or they simply didn't work for some reason. \n\nTo keep this topic rather dense lets avoid what's already mentioned in other threads (ensemble, pseudolabels, etc.) \n\nFor instance optimizing [kappa loss](https://www.kaggle.com/zaharch/minimizing-qwk-directly-with-nn) directly didn't work for me (2x runs with 2 folds each)\n\n[UPDATE] here I provide the list of tricks mentioned in this thread\n- Optimize kappa loss directly\n- Attention Network - [CBAM](https://eccv2018.org/openaccess/content_ECCV_2018/papers/Sanghyun_Woo_Convolutional_Block_Attention_ECCV_2018_paper.pdf)",
      "votes": null
    },
    {
      "id": "622851",
      "postDate": "09/10/2019 06:50:30",
      "content": "<p>I wish I had tried one of those fancy attention networks, I'm not much familiar with those tbh. </p>",
      "rawMarkdown": "I wish I had tried one of those fancy attention networks, I'm not much familiar with those tbh.",
      "votes": null
    },
    {
      "id": "622868",
      "postDate": "09/10/2019 07:21:31",
      "content": "<p>I had it on my list but didn't do it either.</p>",
      "rawMarkdown": "I had it on my list but didn't do it either.",
      "votes": null
    },
    {
      "id": "622952",
      "postDate": "09/10/2019 09:51:59",
      "content": "<p>I wish I had done model ensemble and TTA correctly, it gave me a big boost after I figured it out. But the competition is already over so yeah .-.</p>",
      "rawMarkdown": "I wish I had done model ensemble and TTA correctly, it gave me a big boost after I figured it out. But the competition is already over so yeah .-.",
      "votes": null
    },
    {
      "id": "622982",
      "postDate": "09/10/2019 10:37:12",
      "content": "<p>I tried basic versions without any help. There might be more to it though.</p>",
      "rawMarkdown": "I tried basic versions without any help. There might be more to it though.",
      "votes": null
    },
    {
      "id": "623703",
      "postDate": "09/11/2019 08:19:25",
      "content": "<p>what is the meaning of fancy attention networks?</p>",
      "rawMarkdown": "what is the meaning of fancy attention networks?",
      "votes": null
    },
    {
      "id": "623743",
      "postDate": "09/11/2019 09:00:30",
      "content": "<p>to focus on important features, it also helps with noise. Bestfitting used it in SIIM :)</p>",
      "rawMarkdown": "to focus on important features, it also helps with noise. Bestfitting used it in SIIM :)",
      "votes": null
    },
    {
      "id": "623750",
      "postDate": "09/11/2019 09:08:54",
      "content": "<p>this was already mentioned elsewhere and could be find in kernels too.</p>",
      "rawMarkdown": "this was already mentioned elsewhere and could be find in kernels too.",
      "votes": null
    },
    {
      "id": "624399",
      "postDate": "09/12/2019 03:14:11",
      "content": "<p>and a lot of top solutions for CHAMPS utilized attention as well</p>",
      "rawMarkdown": "and a lot of top solutions for CHAMPS utilized attention as well",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 622851,
      "author_name": "rishabhiitbhu",
      "author_url": "",
      "post_date": "09/10/2019 06:50:30",
      "content": "<p>I wish I had tried one of those fancy attention networks, I'm not much familiar with those tbh. </p>",
      "votes": null,
      "replies": [
        {
          "id": 622868,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "09/10/2019 07:21:31",
          "content": "<p>I had it on my list but didn't do it either.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 622982,
          "author_name": "philippsinger",
          "author_url": "",
          "post_date": "09/10/2019 10:37:12",
          "content": "<p>I tried basic versions without any help. There might be more to it though.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 623703,
          "author_name": "dldmw579",
          "author_url": "",
          "post_date": "09/11/2019 08:19:25",
          "content": "<p>what is the meaning of fancy attention networks?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 623743,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "09/11/2019 09:00:30",
          "content": "<p>to focus on important features, it also helps with noise. Bestfitting used it in SIIM :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 624399,
          "author_name": "sidhanthholalkere",
          "author_url": "",
          "post_date": "09/12/2019 03:14:11",
          "content": "<p>and a lot of top solutions for CHAMPS utilized attention as well</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 622952,
      "author_name": "quandapro",
      "author_url": "",
      "post_date": "09/10/2019 09:51:59",
      "content": "<p>I wish I had done model ensemble and TTA correctly, it gave me a big boost after I figured it out. But the competition is already over so yeah .-.</p>",
      "votes": null,
      "replies": [
        {
          "id": 623750,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "09/11/2019 09:08:54",
          "content": "<p>this was already mentioned elsewhere and could be find in kernels too.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "622844": "Let's share the ideas you came up with but you haven't had time or skill to try or they simply didn't work for some reason. \n\nTo keep this topic rather dense lets avoid what's already mentioned in other threads (ensemble, pseudolabels, etc.) \n\nFor instance optimizing [kappa loss](https://www.kaggle.com/zaharch/minimizing-qwk-directly-with-nn) directly didn't work for me (2x runs with 2 folds each)\n\n[UPDATE] here I provide the list of tricks mentioned in this thread\n- Optimize kappa loss directly\n- Attention Network - [CBAM](https://eccv2018.org/openaccess/content_ECCV_2018/papers/Sanghyun_Woo_Convolutional_Block_Attention_ECCV_2018_paper.pdf)",
    "622851": "I wish I had tried one of those fancy attention networks, I'm not much familiar with those tbh.",
    "622868": "I had it on my list but didn't do it either.",
    "622952": "I wish I had done model ensemble and TTA correctly, it gave me a big boost after I figured it out. But the competition is already over so yeah .-.",
    "622982": "I tried basic versions without any help. There might be more to it though.",
    "623703": "what is the meaning of fancy attention networks?",
    "623743": "to focus on important features, it also helps with noise. Bestfitting used it in SIIM :)",
    "623750": "this was already mentioned elsewhere and could be find in kernels too.",
    "624399": "and a lot of top solutions for CHAMPS utilized attention as well"
  },
  "source": "meta"
}