{
  "id": 220647,
  "title": "Small margins",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/220647",
  "author_name": "",
  "post_date": "2021-02-19T04:23:01.761368900Z",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Looks like the margins were exceptionally small in this competition. People sharing solution details but looks like there wasn't any great relationship between models and methods and results (except maybe the first few solutions). Maybe a great deal of the complexity is not necessary. </p>\n<p>Curious what the simplest solution is that could get in the .899-.900 range. Did anyone have a straightforward approach that scored high or was the complexity necessary for a good result?</p>",
  "messages": [
    {
      "id": "1209844",
      "postDate": "02/19/2021 04:23:01",
      "content": "<p>Looks like the margins were exceptionally small in this competition. People sharing solution details but looks like there wasn't any great relationship between models and methods and results (except maybe the first few solutions). Maybe a great deal of the complexity is not necessary. </p>\n<p>Curious what the simplest solution is that could get in the .899-.900 range. Did anyone have a straightforward approach that scored high or was the complexity necessary for a good result?</p>",
      "rawMarkdown": "Looks like the margins were exceptionally small in this competition. People sharing solution details but looks like there wasn't any great relationship between models and methods and results (except maybe the first few solutions). Maybe a great deal of the complexity is not necessary. \n\nCurious what the simplest solution is that could get in the .899-.900 range. Did anyone have a straightforward approach that scored high or was the complexity necessary for a good result?",
      "votes": null
    },
    {
      "id": "1209866",
      "postDate": "02/19/2021 04:41:07",
      "content": "<p>our solution was really simple:<br>\n<a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220628\" target=\"_blank\">https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220628</a></p>\n<p>No 2019 dataset, no technique for dealing with noisy labels, just use three simple models (SE-ResNext50, EfficientNet-B7, EfficientNet-B3a) with basic training techniques. As all of us were quite busy, this is all we had time for lol</p>",
      "rawMarkdown": "our solution was really simple:\nhttps://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220628\n\nNo 2019 dataset, no technique for dealing with noisy labels, just use three simple models (SE-ResNext50, EfficientNet-B7, EfficientNet-B3a) with basic training techniques. As all of us were quite busy, this is all we had time for lol",
      "votes": null
    },
    {
      "id": "1209881",
      "postDate": "02/19/2021 04:53:14",
      "content": "<p>I think luck is more important than skill in this competition except first place of course.</p>",
      "rawMarkdown": "I think luck is more important than skill in this competition except first place of course.",
      "votes": null
    },
    {
      "id": "1209903",
      "postDate": "02/19/2021 05:18:31",
      "content": "<p>Just one single model(5-folds).</p>\n<p>p.s. I'm looking at the my submission, and it seems that 0.899~0.900 is possible only by ensemble two models with 0.5.</p>",
      "rawMarkdown": "Just one single model(5-folds).\n\np.s. I'm looking at the my submission, and it seems that 0.899~0.900 is possible only by ensemble two models with 0.5.",
      "votes": null
    },
    {
      "id": "1211387",
      "postDate": "02/20/2021 07:06:25",
      "content": "<p>Not that I selected it (of course!) but the best Priv Lb 0.9004 (public 0.9025) was ViT 5-fold &amp; Resnet 5-fold 0.5 each 3xTTA.<br>\nLoss was a weighted loss SmoothBCEwLogits which seemed to do well, better than Taylor CE for me.<br>\nViT 5-fold only was 0.8990. Priv LB 0.9007 Public LB<br>\nIt is a mystery how one would to decide to select this…  </p>",
      "rawMarkdown": "Not that I selected it (of course!) but the best Priv Lb 0.9004 (public 0.9025) was ViT 5-fold & Resnet 5-fold 0.5 each 3xTTA.\nLoss was a weighted loss SmoothBCEwLogits which seemed to do well, better than Taylor CE for me.\nViT 5-fold only was 0.8990. Priv LB 0.9007 Public LB\nIt is a mystery how one would to decide to select this...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1209866,
      "author_name": "tanlikesmath",
      "author_url": "",
      "post_date": "02/19/2021 04:41:07",
      "content": "<p>our solution was really simple:<br>\n<a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220628\" target=\"_blank\">https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220628</a></p>\n<p>No 2019 dataset, no technique for dealing with noisy labels, just use three simple models (SE-ResNext50, EfficientNet-B7, EfficientNet-B3a) with basic training techniques. As all of us were quite busy, this is all we had time for lol</p>",
      "votes": null,
      "replies": [
        {
          "id": 1209881,
          "author_name": "underwearfitting",
          "author_url": "",
          "post_date": "02/19/2021 04:53:14",
          "content": "<p>I think luck is more important than skill in this competition except first place of course.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1209903,
      "author_name": "piantic",
      "author_url": "",
      "post_date": "02/19/2021 05:18:31",
      "content": "<p>Just one single model(5-folds).</p>\n<p>p.s. I'm looking at the my submission, and it seems that 0.899~0.900 is possible only by ensemble two models with 0.5.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1211387,
          "author_name": "something4kag",
          "author_url": "",
          "post_date": "02/20/2021 07:06:25",
          "content": "<p>Not that I selected it (of course!) but the best Priv Lb 0.9004 (public 0.9025) was ViT 5-fold &amp; Resnet 5-fold 0.5 each 3xTTA.<br>\nLoss was a weighted loss SmoothBCEwLogits which seemed to do well, better than Taylor CE for me.<br>\nViT 5-fold only was 0.8990. Priv LB 0.9007 Public LB<br>\nIt is a mystery how one would to decide to select this…  </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1209844": "Looks like the margins were exceptionally small in this competition. People sharing solution details but looks like there wasn't any great relationship between models and methods and results (except maybe the first few solutions). Maybe a great deal of the complexity is not necessary. \n\nCurious what the simplest solution is that could get in the .899-.900 range. Did anyone have a straightforward approach that scored high or was the complexity necessary for a good result?",
    "1209866": "our solution was really simple:\nhttps://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220628\n\nNo 2019 dataset, no technique for dealing with noisy labels, just use three simple models (SE-ResNext50, EfficientNet-B7, EfficientNet-B3a) with basic training techniques. As all of us were quite busy, this is all we had time for lol",
    "1209881": "I think luck is more important than skill in this competition except first place of course.",
    "1209903": "Just one single model(5-folds).\n\np.s. I'm looking at the my submission, and it seems that 0.899~0.900 is possible only by ensemble two models with 0.5.",
    "1211387": "Not that I selected it (of course!) but the best Priv Lb 0.9004 (public 0.9025) was ViT 5-fold & Resnet 5-fold 0.5 each 3xTTA.\nLoss was a weighted loss SmoothBCEwLogits which seemed to do well, better than Taylor CE for me.\nViT 5-fold only was 0.8990. Priv LB 0.9007 Public LB\nIt is a mystery how one would to decide to select this..."
  },
  "source": "meta"
}