{
  "id": 108097,
  "title": "198th place solution summary",
  "url": "/competitions/aptos2019-blindness-detection/discussion/108097",
  "author_name": "Hilal Shaath",
  "post_date": "2019-09-09T05:06:00.845000",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>This was a very illusive competition, but it was for a wonderful cause, and was organized pretty well.</p>\n\n<p>For me, I used efficient net 7 , and an ensemble of efficient net 6 and 7.  For pre-processing I used Ben Graham's method, I ran out of time after figuring the right batch size and learning rate for efficient 7. Cross fold validation didn't predict the very well in my case, but that's probably for not finding the right fold size (I tried 3 and 5).  I tried different architectures DenseNet(X), Resnet(X), efficient yielded the best results, the image size 299 worked best, I suspect bigger image sizes would yield even better results (unfortunately, I didn't have the hardware for that).  I couldn't rely on quadratic kappa (I think everyone saw the shake up), I used F1-macro score instead guage how well my model is doing. </p>\n\n<p>I didn't use focal loss, it didn't give good results.  The batch size for B6/B7 was key to moving the score a little bit up.  </p>\n\n<p>Didn't use external data for many reasons ( I tried, but there was difference of opinion among the classification type).  One thing I really wanted to try was pseudo labeling, but I ran out of time.</p>\n\n<p>I wish to congratulate the winners, really amazing insight.  I hope we can have more medical related competitions. </p>\n\n<p>And I wish to thank the whole community for publishing amazing Kernels, and great write ups that I really learn a lot from.</p>",
  "messages": [
    {
      "id": 621920,
      "postDate": "2019-09-09T05:06:00.847Z",
      "content": "<p>This was a very illusive competition, but it was for a wonderful cause, and was organized pretty well.</p>\n\n<p>For me, I used efficient net 7 , and an ensemble of efficient net 6 and 7.  For pre-processing I used Ben Graham's method, I ran out of time after figuring the right batch size and learning rate for efficient 7. Cross fold validation didn't predict the very well in my case, but that's probably for not finding the right fold size (I tried 3 and 5).  I tried different architectures DenseNet(X), Resnet(X), efficient yielded the best results, the image size 299 worked best, I suspect bigger image sizes would yield even better results (unfortunately, I didn't have the hardware for that).  I couldn't rely on quadratic kappa (I think everyone saw the shake up), I used F1-macro score instead guage how well my model is doing. </p>\n\n<p>I didn't use focal loss, it didn't give good results.  The batch size for B6/B7 was key to moving the score a little bit up.  </p>\n\n<p>Didn't use external data for many reasons ( I tried, but there was difference of opinion among the classification type).  One thing I really wanted to try was pseudo labeling, but I ran out of time.</p>\n\n<p>I wish to congratulate the winners, really amazing insight.  I hope we can have more medical related competitions. </p>\n\n<p>And I wish to thank the whole community for publishing amazing Kernels, and great write ups that I really learn a lot from.</p>",
      "rawMarkdown": "This was a very illusive competition, but it was for a wonderful cause, and was organized pretty well.\n\nFor me, I used efficient net 7 , and an ensemble of efficient net 6 and 7.  For pre-processing I used Ben Graham's method, I ran out of time after figuring the right batch size and learning rate for efficient 7. Cross fold validation didn't predict the very well in my case, but that's probably for not finding the right fold size (I tried 3 and 5).  I tried different architectures DenseNet(X), Resnet(X), efficient yielded the best results, the image size 299 worked best, I suspect bigger image sizes would yield even better results (unfortunately, I didn't have the hardware for that).  I couldn't rely on quadratic kappa (I think everyone saw the shake up), I used F1-macro score instead guage how well my model is doing. \n\nI didn't use focal loss, it didn't give good results.  The batch size for B6/B7 was key to moving the score a little bit up.  \n\nDidn't use external data for many reasons ( I tried, but there was difference of opinion among the classification type).  One thing I really wanted to try was pseudo labeling, but I ran out of time.\n\nI wish to congratulate the winners, really amazing insight.  I hope we can have more medical related competitions. \n\nAnd I wish to thank the whole community for publishing amazing Kernels, and great write ups that I really learn a lot from.",
      "votes": 4
    },
    {
      "id": 623460,
      "postDate": "2019-09-11T01:49:01.087Z",
      "content": "<p>Thank you <a href=\"/veeralakrishna\">@veeralakrishna</a>  for your kind words.</p>",
      "rawMarkdown": "Thank you @veeralakrishna  for your kind words."
    },
    {
      "id": 621956,
      "postDate": "2019-09-09T05:51:51.147Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 623460,
      "author_name": "Hilal Shaath",
      "author_url": "",
      "post_date": "2019-09-11T01:49:01.087000",
      "content": "<p>Thank you <a href=\"/veeralakrishna\">@veeralakrishna</a>  for your kind words.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 621956,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-09-09T05:51:51.147000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "621920": "This was a very illusive competition, but it was for a wonderful cause, and was organized pretty well.\n\nFor me, I used efficient net 7 , and an ensemble of efficient net 6 and 7.  For pre-processing I used Ben Graham's method, I ran out of time after figuring the right batch size and learning rate for efficient 7. Cross fold validation didn't predict the very well in my case, but that's probably for not finding the right fold size (I tried 3 and 5).  I tried different architectures DenseNet(X), Resnet(X), efficient yielded the best results, the image size 299 worked best, I suspect bigger image sizes would yield even better results (unfortunately, I didn't have the hardware for that).  I couldn't rely on quadratic kappa (I think everyone saw the shake up), I used F1-macro score instead guage how well my model is doing. \n\nI didn't use focal loss, it didn't give good results.  The batch size for B6/B7 was key to moving the score a little bit up.  \n\nDidn't use external data for many reasons ( I tried, but there was difference of opinion among the classification type).  One thing I really wanted to try was pseudo labeling, but I ran out of time.\n\nI wish to congratulate the winners, really amazing insight.  I hope we can have more medical related competitions. \n\nAnd I wish to thank the whole community for publishing amazing Kernels, and great write ups that I really learn a lot from.",
    "623460": "Thank you @veeralakrishna  for your kind words.",
    "621956": ""
  }
}