{
  "id": 112788,
  "title": "1st place solution",
  "url": "/competitions/kuzushiji-recognition/discussion/112788",
  "author_name": "tascj",
  "post_date": "2019-10-15T11:47:00.914000",
  "votes": 65,
  "comment_count": 12,
  "views": 0,
  "content": "<p>Thanks for this interesting challenge, and congrats to winners.</p>\n\n<p>My solution is straightfoward. Not much beyond a detection baseline.</p>\n\n<p>Cascade R-CNN with:\n* Strong backbones\n* Multi-scale train&amp;test</p>\n\n<p>Due to limited GPU memory, models were trained on 1024x1024 crops and tested on full images(with a max size limit).</p>\n\n<p>LB score 0.935 with:\n* HRNet w32\n* train scales 512~768\n* test scales [0.5, 0.625, 0.75]</p>\n\n<p>LB score 0.946 with:\n* HRNet w32\n* train scales 768~1280\n* test scales [0.75, 0.875, 1.0, 1.125, 1.25]</p>\n\n<p>Ensembling HRNet_w32 and HRNet_w48 results -&gt; 0.950.</p>\n\n<p>Code: <a href=\"https://github.com/tascj/kaggle-kuzushiji-recognition\">https://github.com/tascj/kaggle-kuzushiji-recognition</a></p>\n\n<p>Trained weights: <a href=\"https://github.com/tascj/kaggle-kuzushiji-recognition/releases\">https://github.com/tascj/kaggle-kuzushiji-recognition/releases</a></p>",
  "messages": [
    {
      "id": 649466,
      "postDate": "2019-10-15T11:47:00.913Z",
      "content": "<p>Thanks for this interesting challenge, and congrats to winners.</p>\n\n<p>My solution is straightfoward. Not much beyond a detection baseline.</p>\n\n<p>Cascade R-CNN with:\n* Strong backbones\n* Multi-scale train&amp;test</p>\n\n<p>Due to limited GPU memory, models were trained on 1024x1024 crops and tested on full images(with a max size limit).</p>\n\n<p>LB score 0.935 with:\n* HRNet w32\n* train scales 512~768\n* test scales [0.5, 0.625, 0.75]</p>\n\n<p>LB score 0.946 with:\n* HRNet w32\n* train scales 768~1280\n* test scales [0.75, 0.875, 1.0, 1.125, 1.25]</p>\n\n<p>Ensembling HRNet_w32 and HRNet_w48 results -&gt; 0.950.</p>\n\n<p>Code: <a href=\"https://github.com/tascj/kaggle-kuzushiji-recognition\">https://github.com/tascj/kaggle-kuzushiji-recognition</a></p>\n\n<p>Trained weights: <a href=\"https://github.com/tascj/kaggle-kuzushiji-recognition/releases\">https://github.com/tascj/kaggle-kuzushiji-recognition/releases</a></p>",
      "rawMarkdown": "Thanks for this interesting challenge, and congrats to winners.\n\nMy solution is straightfoward. Not much beyond a detection baseline.\n\nCascade R-CNN with:\n* Strong backbones\n* Multi-scale train&amp;test\n\nDue to limited GPU memory, models were trained on 1024x1024 crops and tested on full images(with a max size limit).\n\nLB score 0.935 with:\n* HRNet w32\n* train scales 512~768\n* test scales [0.5, 0.625, 0.75]\n\nLB score 0.946 with:\n* HRNet w32\n* train scales 768~1280\n* test scales [0.75, 0.875, 1.0, 1.125, 1.25]\n\nEnsembling HRNet\\_w32 and HRNet\\_w48 results -&gt; 0.950.\n\n\nCode: https://github.com/tascj/kaggle-kuzushiji-recognition\n\nTrained weights: https://github.com/tascj/kaggle-kuzushiji-recognition/releases",
      "votes": 65
    },
    {
      "id": 649603,
      "postDate": "2019-10-15T15:07:25.437Z",
      "content": "<p>Thank you for sharing the solution and congrats! 🎉 </p>",
      "rawMarkdown": "Thank you for sharing the solution and congrats! 🎉 ",
      "votes": 1
    },
    {
      "id": 649680,
      "postDate": "2019-10-15T16:46:39.197Z",
      "content": "<p>Thank you so much and congratulations!</p>",
      "rawMarkdown": "Thank you so much and congratulations!",
      "votes": 2
    },
    {
      "id": 649594,
      "postDate": "2019-10-15T14:57:11.733Z",
      "content": "<p>Congratulations on Winning the Competition\nGreat Write-Up\nThanks for Sharing your Code &amp; Insights.. <a href=\"/tascj0\">@tascj0</a> </p>",
      "rawMarkdown": "Congratulations on Winning the Competition\nGreat Write-Up\nThanks for Sharing your Code &amp; Insights.. @tascj0 ",
      "votes": 2
    },
    {
      "id": 649582,
      "postDate": "2019-10-15T14:46:35.803Z",
      "content": "<p>Congratulations, thanks for sharing your solution!</p>",
      "rawMarkdown": "Congratulations, thanks for sharing your solution!",
      "votes": 2
    },
    {
      "id": 769292,
      "postDate": "2020-03-11T18:28:44.893Z",
      "content": "<p>Hi excellent contribution, is there any general way to get an architecture in the for of the configurations required by mmdetection?</p>",
      "rawMarkdown": "Hi excellent contribution, is there any general way to get an architecture in the for of the configurations required by mmdetection?"
    },
    {
      "id": 675394,
      "postDate": "2019-11-18T03:04:48.310Z",
      "content": "<p>Hello, </p>\n\n<p>Organizer here.  If you wouldn't mind, can you say how much time your method takes for inference per-page-image (even a ballpark or rough estimate is fine)?  </p>\n\n<p>Best, </p>\n\n<p>Alex.  </p>",
      "rawMarkdown": "Hello, \n\nOrganizer here.  If you wouldn't mind, can you say how much time your method takes for inference per-page-image (even a ballpark or rough estimate is fine)?  \n\nBest, \n\nAlex.  ",
      "replies": [
        {
          "id": 675758,
          "postDate": "2019-11-18T14:16:32.410Z",
          "content": "<p>A quick test,\n- 1080 Ti\n- HRNetW32 backbone\n- single scale (0.75) test\n- LB score 0.941/0.942\n- around 0.8s per page</p>\n\n<p>I guess multi-scale test and ensemble would take 20~30x more time.</p>",
          "rawMarkdown": "A quick test,\n- 1080 Ti\n- HRNetW32 backbone\n- single scale (0.75) test\n- LB score 0.941/0.942\n- around 0.8s per page\n\nI guess multi-scale test and ensemble would take 20~30x more time.",
          "votes": 1
        }
      ]
    },
    {
      "id": 656541,
      "postDate": "2019-10-24T11:11:40.050Z",
      "content": "<p>Thank you for sharing and congrats! It is very interesting.</p>",
      "rawMarkdown": "Thank you for sharing and congrats! It is very interesting.\n"
    },
    {
      "id": 1092635,
      "postDate": "2020-11-27T04:07:09.083Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 650694,
      "postDate": "2019-10-16T15:49:56.163Z",
      "content": "<p>Thank you very much!!!</p>",
      "rawMarkdown": "Thank you very much!!!",
      "votes": 1
    },
    {
      "id": 668178,
      "postDate": "2019-11-08T04:57:20.767Z",
      "content": "<p>Thank you for sharing and congrats!</p>",
      "rawMarkdown": "Thank you for sharing and congrats!"
    },
    {
      "id": 664727,
      "postDate": "2019-11-04T06:20:14.047Z",
      "content": "<p>Thanks for sharing, bro!</p>",
      "rawMarkdown": "Thanks for sharing, bro!"
    }
  ],
  "comments": [
    {
      "id": 649603,
      "author_name": "Tanyapohn",
      "author_url": "",
      "post_date": "2019-10-15T15:07:25.437000",
      "content": "<p>Thank you for sharing the solution and congrats! 🎉 </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 649680,
      "author_name": "tkasasagi",
      "author_url": "",
      "post_date": "2019-10-15T16:46:39.197000",
      "content": "<p>Thank you so much and congratulations!</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 649594,
      "author_name": "Ailurophile",
      "author_url": "",
      "post_date": "2019-10-15T14:57:11.733000",
      "content": "<p>Congratulations on Winning the Competition\nGreat Write-Up\nThanks for Sharing your Code &amp; Insights.. <a href=\"/tascj0\">@tascj0</a> </p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 649582,
      "author_name": "Anna Novikova",
      "author_url": "",
      "post_date": "2019-10-15T14:46:35.803000",
      "content": "<p>Congratulations, thanks for sharing your solution!</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 769292,
      "author_name": "Felipe Bivort Haiek",
      "author_url": "",
      "post_date": "2020-03-11T18:28:44.893000",
      "content": "<p>Hi excellent contribution, is there any general way to get an architecture in the for of the configurations required by mmdetection?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 675394,
      "author_name": "TheNuttyNetter",
      "author_url": "",
      "post_date": "2019-11-18T03:04:48.310000",
      "content": "<p>Hello, </p>\n\n<p>Organizer here.  If you wouldn't mind, can you say how much time your method takes for inference per-page-image (even a ballpark or rough estimate is fine)?  </p>\n\n<p>Best, </p>\n\n<p>Alex.  </p>",
      "votes": 0,
      "replies": [
        {
          "id": 675758,
          "author_name": "tascj",
          "author_url": "",
          "post_date": "2019-11-18T14:16:32.410000",
          "content": "<p>A quick test,\n- 1080 Ti\n- HRNetW32 backbone\n- single scale (0.75) test\n- LB score 0.941/0.942\n- around 0.8s per page</p>\n\n<p>I guess multi-scale test and ensemble would take 20~30x more time.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 656541,
      "author_name": "Masanori Shimao",
      "author_url": "",
      "post_date": "2019-10-24T11:11:40.050000",
      "content": "<p>Thank you for sharing and congrats! It is very interesting.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1092635,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-11-27T04:07:09.083000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 650694,
      "author_name": "Ma Yuxuan",
      "author_url": "",
      "post_date": "2019-10-16T15:49:56.163000",
      "content": "<p>Thank you very much!!!</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 668178,
      "author_name": "sjsjs",
      "author_url": "",
      "post_date": "2019-11-08T04:57:20.767000",
      "content": "<p>Thank you for sharing and congrats!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 664727,
      "author_name": "Nguyễn Duy Cương",
      "author_url": "",
      "post_date": "2019-11-04T06:20:14.047000",
      "content": "<p>Thanks for sharing, bro!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "649466": "Thanks for this interesting challenge, and congrats to winners.\n\nMy solution is straightfoward. Not much beyond a detection baseline.\n\nCascade R-CNN with:\n* Strong backbones\n* Multi-scale train&amp;test\n\nDue to limited GPU memory, models were trained on 1024x1024 crops and tested on full images(with a max size limit).\n\nLB score 0.935 with:\n* HRNet w32\n* train scales 512~768\n* test scales [0.5, 0.625, 0.75]\n\nLB score 0.946 with:\n* HRNet w32\n* train scales 768~1280\n* test scales [0.75, 0.875, 1.0, 1.125, 1.25]\n\nEnsembling HRNet\\_w32 and HRNet\\_w48 results -&gt; 0.950.\n\n\nCode: https://github.com/tascj/kaggle-kuzushiji-recognition\n\nTrained weights: https://github.com/tascj/kaggle-kuzushiji-recognition/releases",
    "649603": "Thank you for sharing the solution and congrats! 🎉 ",
    "649680": "Thank you so much and congratulations!",
    "649594": "Congratulations on Winning the Competition\nGreat Write-Up\nThanks for Sharing your Code &amp; Insights.. @tascj0 ",
    "649582": "Congratulations, thanks for sharing your solution!",
    "769292": "Hi excellent contribution, is there any general way to get an architecture in the for of the configurations required by mmdetection?",
    "675394": "Hello, \n\nOrganizer here.  If you wouldn't mind, can you say how much time your method takes for inference per-page-image (even a ballpark or rough estimate is fine)?  \n\nBest, \n\nAlex.  ",
    "656541": "Thank you for sharing and congrats! It is very interesting.\n",
    "1092635": "",
    "650694": "Thank you very much!!!",
    "668178": "Thank you for sharing and congrats!",
    "664727": "Thanks for sharing, bro!"
  }
}