{
  "id": 71617,
  "title": "Significant shakeup on private leaderboard",
  "url": "/competitions/airbus-ship-detection/discussion/71617",
  "author_name": "",
  "post_date": "2018-11-15T06:51:41.004197900Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>First, Congratulate to all winners !</p>\n\n<p>Many thank @lafoss to share your kernel and thought on this competitions, I learned a lot from you and your comment.</p>\n\n<p>I was really frustrated that got worse private grade than LB grade ….\nI finally beat the 0.699 LB score that from lafoss <a href=\"https://www.kaggle.com/iafoss/unet34-submission-tta-0-699-new-public-lb\">kernel</a>, but get the worse private score.\nand the grade suddenly drop a lot …. This is my first DL competition on Kaggle. Is this situation normally? (I mean have Significant shakeup on private leaderboard )</p>\n\n<p>Want to know why this happened. It looks like overfitting issue, how to prevent it and what is the key point of private score ? I also try split group BigImageID on validation from Emil's discussion. but it seems still have this issue. Want to know why and try to prevent it next time. Thanks.</p>\n\n<p>I found one thing on my submisstions, This is my step to make submit file:\n1.  Train model for predict ship or not\n2. ResNet34+Unet for image segmentation\n3. Use 1 to predict image have ship or not, only don 2 on image that have ship</p>\n\n<p>I found if I don't use Step1 model to predict  ship exist even if I got better LB score, the private score is worse.... so I think part1 is really important... </p>\n\n<p>Do not trust LB...</p>",
  "messages": [
    {
      "id": "421583",
      "postDate": "11/15/2018 06:51:41",
      "content": "<p>First, Congratulate to all winners !</p>\n\n<p>Many thank @lafoss to share your kernel and thought on this competitions, I learned a lot from you and your comment.</p>\n\n<p>I was really frustrated that got worse private grade than LB grade ….\nI finally beat the 0.699 LB score that from lafoss <a href=\"https://www.kaggle.com/iafoss/unet34-submission-tta-0-699-new-public-lb\">kernel</a>, but get the worse private score.\nand the grade suddenly drop a lot …. This is my first DL competition on Kaggle. Is this situation normally? (I mean have Significant shakeup on private leaderboard )</p>\n\n<p>Want to know why this happened. It looks like overfitting issue, how to prevent it and what is the key point of private score ? I also try split group BigImageID on validation from Emil's discussion. but it seems still have this issue. Want to know why and try to prevent it next time. Thanks.</p>\n\n<p>I found one thing on my submisstions, This is my step to make submit file:\n1.  Train model for predict ship or not\n2. ResNet34+Unet for image segmentation\n3. Use 1 to predict image have ship or not, only don 2 on image that have ship</p>\n\n<p>I found if I don't use Step1 model to predict  ship exist even if I got better LB score, the private score is worse.... so I think part1 is really important... </p>\n\n<p>Do not trust LB...</p>",
      "rawMarkdown": "First, Congratulate to all winners !\n\nMany thank @lafoss to share your kernel and thought on this competitions, I learned a lot from you and your comment.\n\nI was really frustrated that got worse private grade than LB grade ….\nI finally beat the 0.699 LB score that from lafoss [kernel][1], but get the worse private score.\nand the grade suddenly drop a lot …. This is my first DL competition on Kaggle. Is this situation normally? (I mean have Significant shakeup on private leaderboard )\n\nWant to know why this happened. It looks like overfitting issue, how to prevent it and what is the key point of private score ? I also try split group BigImageID on validation from Emil's discussion. but it seems still have this issue. Want to know why and try to prevent it next time. Thanks.\n\nI found one thing on my submisstions, This is my step to make submit file:\n1.  Train model for predict ship or not\n2. ResNet34+Unet for image segmentation\n3. Use 1 to predict image have ship or not, only don 2 on image that have ship\n\nI found if I don't use Step1 model to predict  ship exist even if I got better LB score, the private score is worse.... so I think part1 is really important... \n\n [1]: https://www.kaggle.com/iafoss/unet34-submission-tta-0-699-new-public-lb\n\nDo not trust LB...",
      "votes": null
    },
    {
      "id": "421713",
      "postDate": "11/15/2018 09:59:49",
      "content": "<p>Yes, huge thanks to @lafoss !\nI'm in the top 100 only by training his Kernels as far as I could.\nI'm sorry to see that he loose so many place in the private LB.</p>",
      "rawMarkdown": "Yes, huge thanks to @lafoss !\nI'm in the top 100 only by training his Kernels as far as I could.\nI'm sorry to see that he loose so many place in the private LB.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 421713,
      "author_name": "trancept",
      "author_url": "",
      "post_date": "11/15/2018 09:59:49",
      "content": "<p>Yes, huge thanks to @lafoss !\nI'm in the top 100 only by training his Kernels as far as I could.\nI'm sorry to see that he loose so many place in the private LB.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "421583": "First, Congratulate to all winners !\n\nMany thank @lafoss to share your kernel and thought on this competitions, I learned a lot from you and your comment.\n\nI was really frustrated that got worse private grade than LB grade ….\nI finally beat the 0.699 LB score that from lafoss [kernel][1], but get the worse private score.\nand the grade suddenly drop a lot …. This is my first DL competition on Kaggle. Is this situation normally? (I mean have Significant shakeup on private leaderboard )\n\nWant to know why this happened. It looks like overfitting issue, how to prevent it and what is the key point of private score ? I also try split group BigImageID on validation from Emil's discussion. but it seems still have this issue. Want to know why and try to prevent it next time. Thanks.\n\nI found one thing on my submisstions, This is my step to make submit file:\n1.  Train model for predict ship or not\n2. ResNet34+Unet for image segmentation\n3. Use 1 to predict image have ship or not, only don 2 on image that have ship\n\nI found if I don't use Step1 model to predict  ship exist even if I got better LB score, the private score is worse.... so I think part1 is really important... \n\n [1]: https://www.kaggle.com/iafoss/unet34-submission-tta-0-699-new-public-lb\n\nDo not trust LB...",
    "421713": "Yes, huge thanks to @lafoss !\nI'm in the top 100 only by training his Kernels as far as I could.\nI'm sorry to see that he loose so many place in the private LB."
  },
  "source": "meta"
}