{
  "id": 77276,
  "title": "25th solution overview",
  "url": "/competitions/human-protein-atlas-image-classification/writeups/soonhwan-kwon-25th-solution-overview",
  "author_name": "",
  "post_date": "2019-01-11T06:06:04.937Z",
  "votes": 15,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Data Processing : simple rotate, flip on 512x512/1024x1024 RGBY</p>\n\n<p>Loss : \nI used focal loss and took lots of time to optimize gamma(I used 2 and 2.5 in final version), and it seems alpha=1, gamma=2.5 actually works better on public leader board but not so good on private leader board.</p>\n\n<p>Model : \nI used  se-resnext50 on 512x512/1024x1024 size images as baseline model and resnet34 on 512 size as a low capacity model. I made 5 cross validation models(total 24 models) and weighted averaged them(I gave more weight on gamma=2.5 models). Due to the limitation of time and resources, I couldn't make the full planned ensemble models.</p>\n\n<p>Thoughts:\nAlthough I exhausted to make 24 models as an ensemble and it seemed worked well in public leader board, but my hidden best private score was from the ensemble which I just 'or'ed all output labels of two single models and have relatively low public score,(0.571/0.538),(0.561/0.525) each(※(public LB/private LB))). And It just turned out that it reached private LB 0.550 which was slightly better than my final ensemble model,0.547. They were from resnet34 and se-resnext50 on both 512 size images, so maybe 24 models for ensemble was too much.</p>",
  "messages": [
    {
      "id": "454041",
      "postDate": "01/11/2019 04:08:25",
      "content": "<p>Data Processing : simple rotate, flip on 512x512/1024x1024 RGBY</p>\n\n<p>Loss : \nI used focal loss and took lots of time to optimize gamma(I used 2 and 2.5 in final version), and it seems alpha=1, gamma=2.5 actually works better on public leader board but not so good on private leader board.</p>\n\n<p>Model : \nI used  se-resnext50 on 512x512/1024x1024 size images as baseline model and resnet34 on 512 size as a low capacity model. I made 5 cross validation models(total 24 models) and weighted averaged them(I gave more weight on gamma=2.5 models). Due to the limitation of time and resources, I couldn't make the full planned ensemble models.</p>\n\n<p>Thoughts:\nAlthough I exhausted to make 24 models as an ensemble and it seemed worked well in public leader board, but my hidden best private score was from the ensemble which I just 'or'ed all output labels of two single models and have relatively low public score,(0.571/0.538),(0.561/0.525) each(※(public LB/private LB))). And It just turned out that it reached private LB 0.550 which was slightly better than my final ensemble model,0.547. They were from resnet34 and se-resnext50 on both 512 size images, so maybe 24 models for ensemble was too much.</p>",
      "rawMarkdown": "Data Processing : simple rotate, flip on 512x512/1024x1024 RGBY\n\nLoss : \nI used focal loss and took lots of time to optimize gamma(I used 2 and 2.5 in final version), and it seems alpha=1, gamma=2.5 actually works better on public leader board but not so good on private leader board.\n\nModel : \nI used  se-resnext50 on 512x512/1024x1024 size images as baseline model and resnet34 on 512 size as a low capacity model. I made 5 cross validation models(total 24 models) and weighted averaged them(I gave more weight on gamma=2.5 models). Due to the limitation of time and resources, I couldn't make the full planned ensemble models.\n\nThoughts:\nAlthough I exhausted to make 24 models as an ensemble and it seemed worked well in public leader board, but my hidden best private score was from the ensemble which I just 'or'ed all output labels of two single models and have relatively low public score,(0.571/0.538),(0.561/0.525) each(※(public LB/private LB))). And It just turned out that it reached private LB 0.550 which was slightly better than my final ensemble model,0.547. They were from resnet34 and se-resnext50 on both 512 size images, so maybe 24 models for ensemble was too much.",
      "votes": null
    },
    {
      "id": "454152",
      "postDate": "01/11/2019 07:11:06",
      "content": "<p>Congrats and thanks!</p>",
      "rawMarkdown": "Congrats and thanks!",
      "votes": null
    },
    {
      "id": "454184",
      "postDate": "01/11/2019 08:10:43",
      "content": "<p>Good job!</p>",
      "rawMarkdown": "Good job!",
      "votes": null
    },
    {
      "id": "454349",
      "postDate": "01/11/2019 13:35:11",
      "content": "<p>congratulation and thanks for sharing..</p>",
      "rawMarkdown": "congratulation and thanks for sharing..",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 454152,
      "author_name": "sgalib",
      "author_url": "",
      "post_date": "01/11/2019 07:11:06",
      "content": "<p>Congrats and thanks!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 454184,
      "author_name": "arnaurm",
      "author_url": "",
      "post_date": "01/11/2019 08:10:43",
      "content": "<p>Good job!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 454349,
      "author_name": "viswanathravindran",
      "author_url": "",
      "post_date": "01/11/2019 13:35:11",
      "content": "<p>congratulation and thanks for sharing..</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "454041": "Data Processing : simple rotate, flip on 512x512/1024x1024 RGBY\n\nLoss : \nI used focal loss and took lots of time to optimize gamma(I used 2 and 2.5 in final version), and it seems alpha=1, gamma=2.5 actually works better on public leader board but not so good on private leader board.\n\nModel : \nI used  se-resnext50 on 512x512/1024x1024 size images as baseline model and resnet34 on 512 size as a low capacity model. I made 5 cross validation models(total 24 models) and weighted averaged them(I gave more weight on gamma=2.5 models). Due to the limitation of time and resources, I couldn't make the full planned ensemble models.\n\nThoughts:\nAlthough I exhausted to make 24 models as an ensemble and it seemed worked well in public leader board, but my hidden best private score was from the ensemble which I just 'or'ed all output labels of two single models and have relatively low public score,(0.571/0.538),(0.561/0.525) each(※(public LB/private LB))). And It just turned out that it reached private LB 0.550 which was slightly better than my final ensemble model,0.547. They were from resnet34 and se-resnext50 on both 512 size images, so maybe 24 models for ensemble was too much.",
    "454152": "Congrats and thanks!",
    "454184": "Good job!",
    "454349": "congratulation and thanks for sharing.."
  },
  "source": "meta"
}