{
  "id": 108143,
  "title": "Private score of 0.923 - Simple Technique",
  "url": "/competitions/aptos2019-blindness-detection/discussion/108143",
  "author_name": "Sterls",
  "post_date": "2019-09-09T12:50:51.085000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Firstly, thanks to @drhabib for sharing his knowledge during this competition. I learnt many new techniques which I'm excited to use in upcoming competitions!  </p>\n\n<p>I was following @drhabib 's comments closely. He mentioned fine-tuning efficientnet b5 which was pre-trained on 2015 and 2019 data. Using this competition's training set, fine tune on image size 224 for 10 epochs, then for larger image sizes, train with 5 epochs. I did that for image sizes 224, 300, 456, and 512. </p>\n\n<p>This simple procedure scored 0.923 on private and 0.787 on public. </p>\n\n<p>This simple procedure, on the private leaderboard, outperformed the models we worked on for weeks. My team and I tried many many things including consultation with domain experts where they advised we try to count the number of microanyeurisms and exudates per image, count the number of hemorrhages per quadrant, etc etc. None of this resulted in a higher private score than @drhabib suggestion. Unfortunately, for the final model, I selected what I thought to be the most robust solution (a stack of efficient nets) which didn't do so well 🤕. </p>\n\n<p>I'll make a notebook over the coming days for the community's help on where we went wrong. </p>",
  "messages": [
    {
      "id": 622260,
      "postDate": "2019-09-09T12:50:51.087Z",
      "content": "<p>Firstly, thanks to @drhabib for sharing his knowledge during this competition. I learnt many new techniques which I'm excited to use in upcoming competitions!  </p>\n\n<p>I was following @drhabib 's comments closely. He mentioned fine-tuning efficientnet b5 which was pre-trained on 2015 and 2019 data. Using this competition's training set, fine tune on image size 224 for 10 epochs, then for larger image sizes, train with 5 epochs. I did that for image sizes 224, 300, 456, and 512. </p>\n\n<p>This simple procedure scored 0.923 on private and 0.787 on public. </p>\n\n<p>This simple procedure, on the private leaderboard, outperformed the models we worked on for weeks. My team and I tried many many things including consultation with domain experts where they advised we try to count the number of microanyeurisms and exudates per image, count the number of hemorrhages per quadrant, etc etc. None of this resulted in a higher private score than @drhabib suggestion. Unfortunately, for the final model, I selected what I thought to be the most robust solution (a stack of efficient nets) which didn't do so well 🤕. </p>\n\n<p>I'll make a notebook over the coming days for the community's help on where we went wrong. </p>",
      "rawMarkdown": "Firstly, thanks to @drhabib for sharing his knowledge during this competition. I learnt many new techniques which I'm excited to use in upcoming competitions!  \n\nI was following @drhabib 's comments closely. He mentioned fine-tuning efficientnet b5 which was pre-trained on 2015 and 2019 data. Using this competition's training set, fine tune on image size 224 for 10 epochs, then for larger image sizes, train with 5 epochs. I did that for image sizes 224, 300, 456, and 512. \n\nThis simple procedure scored 0.923 on private and 0.787 on public. \n\nThis simple procedure, on the private leaderboard, outperformed the models we worked on for weeks. My team and I tried many many things including consultation with domain experts where they advised we try to count the number of microanyeurisms and exudates per image, count the number of hemorrhages per quadrant, etc etc. None of this resulted in a higher private score than @drhabib suggestion. Unfortunately, for the final model, I selected what I thought to be the most robust solution (a stack of efficient nets) which didn't do so well 🤕. \n\nI'll make a notebook over the coming days for the community's help on where we went wrong. ",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "622260": "Firstly, thanks to @drhabib for sharing his knowledge during this competition. I learnt many new techniques which I'm excited to use in upcoming competitions!  \n\nI was following @drhabib 's comments closely. He mentioned fine-tuning efficientnet b5 which was pre-trained on 2015 and 2019 data. Using this competition's training set, fine tune on image size 224 for 10 epochs, then for larger image sizes, train with 5 epochs. I did that for image sizes 224, 300, 456, and 512. \n\nThis simple procedure scored 0.923 on private and 0.787 on public. \n\nThis simple procedure, on the private leaderboard, outperformed the models we worked on for weeks. My team and I tried many many things including consultation with domain experts where they advised we try to count the number of microanyeurisms and exudates per image, count the number of hemorrhages per quadrant, etc etc. None of this resulted in a higher private score than @drhabib suggestion. Unfortunately, for the final model, I selected what I thought to be the most robust solution (a stack of efficient nets) which didn't do so well 🤕. \n\nI'll make a notebook over the coming days for the community's help on where we went wrong. "
  }
}