{
  "id": 98859,
  "title": "Anyone used past competition dataset successfully?",
  "url": "/competitions/aptos2019-blindness-detection/discussion/98859",
  "author_name": "",
  "post_date": "2019-07-07T03:55:53.861862100Z",
  "votes": 4,
  "comment_count": 7,
  "views": 0,
  "content": "<p>There was a similar <a href=\"https://www.kaggle.com/c/diabetic-retinopathy-detection\">competition</a> 4 years ago.\nAnyone used it to pretrain model &amp; use that on this dataset?</p>\n\n<p>I tried using it. But for me, its not helpful at all. Score decreases if I use that data.\nI was using this <a href=\"https://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized\">dataset</a> which is resized dataset of previous competition.</p>",
  "messages": [
    {
      "id": "569631",
      "postDate": "07/07/2019 03:55:53",
      "content": "<p>There was a similar <a href=\"https://www.kaggle.com/c/diabetic-retinopathy-detection\">competition</a> 4 years ago.\nAnyone used it to pretrain model &amp; use that on this dataset?</p>\n\n<p>I tried using it. But for me, its not helpful at all. Score decreases if I use that data.\nI was using this <a href=\"https://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized\">dataset</a> which is resized dataset of previous competition.</p>",
      "rawMarkdown": "There was a similar [competition](https://www.kaggle.com/c/diabetic-retinopathy-detection) 4 years ago.\nAnyone used it to pretrain model &amp; use that on this dataset?\n\nI tried using it. But for me, its not helpful at all. Score decreases if I use that data.\nI was using this [dataset](https://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized) which is resized dataset of previous competition.",
      "votes": null
    },
    {
      "id": "569654",
      "postDate": "07/07/2019 05:29:53",
      "content": "<p>I have trained a pretrained model with the same dataset as you showed, for my classification model, score increased from 0.743 to 0.747;  and for regression model, I only tested model with pretrained model, lb score is 0.782. </p>",
      "rawMarkdown": "I have trained a pretrained model with the same dataset as you showed, for my classification model, score increased from 0.743 to 0.747;  and for regression model, I only tested model with pretrained model, lb score is 0.782.",
      "votes": null
    },
    {
      "id": "569770",
      "postDate": "07/07/2019 09:37:23",
      "content": "<p>Wow, that's very good improvement. I must have committed some mistake implementing it.\nCan you please tell us, what type architecture you're using?</p>",
      "rawMarkdown": "Wow, that's very good improvement. I must have committed some mistake implementing it.\nCan you please tell us, what type architecture you're using?",
      "votes": null
    },
    {
      "id": "569944",
      "postDate": "07/07/2019 15:29:24",
      "content": "<p>seresnext101. (I usually use this first</p>",
      "rawMarkdown": "seresnext101. (I usually use this first",
      "votes": null
    },
    {
      "id": "570285",
      "postDate": "07/08/2019 05:09:00",
      "content": "<p>Thanks for sharing <a href=\"/jionie\">@jionie</a>.\nHow you generally choose # of epochs in training? Based on validation loss or some other method?</p>\n\n<p>What I am doing is, selecting 10 epochs &amp; do augmentation (random horizontal flip &amp; random 360 degree rotate) so it will not overfit on trainset.</p>",
      "rawMarkdown": "Thanks for sharing @jionie.\nHow you generally choose # of epochs in training? Based on validation loss or some other method?\n\nWhat I am doing is, selecting 10 epochs &amp; do augmentation (random horizontal flip &amp; random 360 degree rotate) so it will not overfit on trainset.",
      "votes": null
    },
    {
      "id": "570302",
      "postDate": "07/08/2019 05:29:44",
      "content": "<p>I always use the metric (quadratic weighted kappa here) to choose models to save. </p>",
      "rawMarkdown": "I always use the metric (quadratic weighted kappa here) to choose models to save.",
      "votes": null
    },
    {
      "id": "570360",
      "postDate": "07/08/2019 07:27:16",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "570361",
      "postDate": "07/08/2019 07:28:21",
      "content": "<p>Early Stopping</p>",
      "rawMarkdown": "Early Stopping",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 569654,
      "author_name": "jionie",
      "author_url": "",
      "post_date": "07/07/2019 05:29:53",
      "content": "<p>I have trained a pretrained model with the same dataset as you showed, for my classification model, score increased from 0.743 to 0.747;  and for regression model, I only tested model with pretrained model, lb score is 0.782. </p>",
      "votes": null,
      "replies": [
        {
          "id": 569770,
          "author_name": "prashantkikani",
          "author_url": "",
          "post_date": "07/07/2019 09:37:23",
          "content": "<p>Wow, that's very good improvement. I must have committed some mistake implementing it.\nCan you please tell us, what type architecture you're using?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 569944,
          "author_name": "jionie",
          "author_url": "",
          "post_date": "07/07/2019 15:29:24",
          "content": "<p>seresnext101. (I usually use this first</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 570285,
          "author_name": "prashantkikani",
          "author_url": "",
          "post_date": "07/08/2019 05:09:00",
          "content": "<p>Thanks for sharing <a href=\"/jionie\">@jionie</a>.\nHow you generally choose # of epochs in training? Based on validation loss or some other method?</p>\n\n<p>What I am doing is, selecting 10 epochs &amp; do augmentation (random horizontal flip &amp; random 360 degree rotate) so it will not overfit on trainset.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 570302,
          "author_name": "jionie",
          "author_url": "",
          "post_date": "07/08/2019 05:29:44",
          "content": "<p>I always use the metric (quadratic weighted kappa here) to choose models to save. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 570360,
          "author_name": "nitin29",
          "author_url": "",
          "post_date": "07/08/2019 07:27:16",
          "content": "",
          "votes": null,
          "replies": []
        },
        {
          "id": 570361,
          "author_name": "nitin29",
          "author_url": "",
          "post_date": "07/08/2019 07:28:21",
          "content": "<p>Early Stopping</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "569631": "There was a similar [competition](https://www.kaggle.com/c/diabetic-retinopathy-detection) 4 years ago.\nAnyone used it to pretrain model &amp; use that on this dataset?\n\nI tried using it. But for me, its not helpful at all. Score decreases if I use that data.\nI was using this [dataset](https://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized) which is resized dataset of previous competition.",
    "569654": "I have trained a pretrained model with the same dataset as you showed, for my classification model, score increased from 0.743 to 0.747;  and for regression model, I only tested model with pretrained model, lb score is 0.782.",
    "569770": "Wow, that's very good improvement. I must have committed some mistake implementing it.\nCan you please tell us, what type architecture you're using?",
    "569944": "seresnext101. (I usually use this first",
    "570285": "Thanks for sharing @jionie.\nHow you generally choose # of epochs in training? Based on validation loss or some other method?\n\nWhat I am doing is, selecting 10 epochs &amp; do augmentation (random horizontal flip &amp; random 360 degree rotate) so it will not overfit on trainset.",
    "570302": "I always use the metric (quadratic weighted kappa here) to choose models to save.",
    "570360": "",
    "570361": "Early Stopping"
  },
  "source": "meta"
}