{
  "id": 221739,
  "title": "30th solution summary",
  "url": "/competitions/cassava-leaf-disease-classification/writeups/kazukim-30th-solution-summary",
  "author_name": "",
  "post_date": "2021-02-24T03:06:03.436389500Z",
  "votes": 18,
  "comment_count": 10,
  "views": 0,
  "content": "<h2>Acknowledgements</h2>\n<p>Thanks to Kaggle and hosts for holding this competition.<br>\n   Thanks Kaggler!! This is my first silver medal :) </p>\n<h2>Result</h2>\n<ul>\n<li>30th</li>\n<li>Private Score 0.9010</li>\n<li>Public Score 0.9018</li>\n</ul>\n<h2>Summary</h2>\n<pre><code>My final submission is ensemble of EfficientNetB5 with Noisy Student and ResNext50(32x4d).\n</code></pre>\n<h2>Datasets</h2>\n<pre><code>This competition datasets only. (I don't use 2019 datasets.)\n</code></pre>\n<h2>Preprocessing</h2>\n<pre><code>Image size 512\nAugumentations\n - ResizeCrop\n - Horizontal,VerticalFlip\n - OneOf(RandomBrightness, RandomContrast)\n - OneOf (RandomBlur,MedianBlur,GaussianBlur)\n - ShiftScaleRotate\n - Normalize\n</code></pre>\n<h2>Training</h2>\n<pre><code>Optimizer Adam\nLoss function CrossEntropyLoss\n</code></pre>\n<h2>Base Model CV</h2>\n<pre><code>CV is StratifiedKFold 5 folds.\n</code></pre>\n<ul>\n<li>EfficientNetB5<ul>\n<li>CV average 0.8898</li>\n<li>Public 0.894</li>\n<li>Private 0.897</li></ul></li>\n<li>ResNext50(32x4d)<ul>\n<li>CV average 0.8898</li>\n<li>Public 0.896</li>\n<li>Private 0.894</li></ul></li>\n</ul>\n<p><a href=\"https://github.com/kazukim10/kaggle_CasavaLeafDeseaseClassfication\" target=\"_blank\">https://github.com/kazukim10/kaggle_CasavaLeafDeseaseClassfication</a></p>",
  "messages": [
    {
      "id": "1215898",
      "postDate": "02/24/2021 03:06:03",
      "content": "<h2>Acknowledgements</h2>\n<p>Thanks to Kaggle and hosts for holding this competition.<br>\n   Thanks Kaggler!! This is my first silver medal :) </p>\n<h2>Result</h2>\n<ul>\n<li>30th</li>\n<li>Private Score 0.9010</li>\n<li>Public Score 0.9018</li>\n</ul>\n<h2>Summary</h2>\n<pre><code>My final submission is ensemble of EfficientNetB5 with Noisy Student and ResNext50(32x4d).\n</code></pre>\n<h2>Datasets</h2>\n<pre><code>This competition datasets only. (I don't use 2019 datasets.)\n</code></pre>\n<h2>Preprocessing</h2>\n<pre><code>Image size 512\nAugumentations\n - ResizeCrop\n - Horizontal,VerticalFlip\n - OneOf(RandomBrightness, RandomContrast)\n - OneOf (RandomBlur,MedianBlur,GaussianBlur)\n - ShiftScaleRotate\n - Normalize\n</code></pre>\n<h2>Training</h2>\n<pre><code>Optimizer Adam\nLoss function CrossEntropyLoss\n</code></pre>\n<h2>Base Model CV</h2>\n<pre><code>CV is StratifiedKFold 5 folds.\n</code></pre>\n<ul>\n<li>EfficientNetB5<ul>\n<li>CV average 0.8898</li>\n<li>Public 0.894</li>\n<li>Private 0.897</li></ul></li>\n<li>ResNext50(32x4d)<ul>\n<li>CV average 0.8898</li>\n<li>Public 0.896</li>\n<li>Private 0.894</li></ul></li>\n</ul>\n<p><a href=\"https://github.com/kazukim10/kaggle_CasavaLeafDeseaseClassfication\" target=\"_blank\">https://github.com/kazukim10/kaggle_CasavaLeafDeseaseClassfication</a></p>",
      "rawMarkdown": "## Acknowledgements\n   Thanks to Kaggle and hosts for holding this competition.\n   Thanks Kaggler!! This is my first silver medal :) \n\n## Result\n   - 30th\n   - Private Score 0.9010\n   - Public Score 0.9018\n\n## Summary\n    My final submission is ensemble of EfficientNetB5 with Noisy Student and ResNext50(32x4d).\n\n## Datasets\n    This competition datasets only. (I don't use 2019 datasets.)\n\n## Preprocessing\n    Image size 512\n    Augumentations\n     - ResizeCrop\n     - Horizontal,VerticalFlip\n     - OneOf(RandomBrightness, RandomContrast)\n     - OneOf (RandomBlur,MedianBlur,GaussianBlur)\n     - ShiftScaleRotate\n     - Normalize\n\n## Training\n    Optimizer Adam\n    Loss function CrossEntropyLoss\n\n## Base Model CV\n    CV is StratifiedKFold 5 folds.\n- EfficientNetB5\n    - CV average 0.8898\n    - Public 0.894\n    - Private 0.897\n- ResNext50(32x4d)\n    - CV average 0.8898\n    - Public 0.896\n    - Private 0.894\n\nhttps://github.com/kazukim10/kaggle_CasavaLeafDeseaseClassfication",
      "votes": null
    },
    {
      "id": "1215939",
      "postDate": "02/24/2021 04:17:46",
      "content": "<p>This model doesn't have that much tricks. Just use CrossEntropyLoss And Neither rigid regularization/generalization technique like TTA, Augmentation, nor Additional Extra Datasets.</p>\n<p>Today I learn,  Simple is The Best.</p>",
      "rawMarkdown": "This model doesn't have that much tricks. Just use CrossEntropyLoss And Neither rigid regularization/generalization technique like TTA, Augmentation, nor Additional Extra Datasets.\n\nToday I learn,  Simple is The Best.",
      "votes": null
    },
    {
      "id": "1215970",
      "postDate": "02/24/2021 04:28:45",
      "content": "<p>Nice job! Simple is always best. <a href=\"https://www.kaggle.com/kazukim\" target=\"_blank\">@kazukim</a> </p>",
      "rawMarkdown": "Nice job! Simple is always best. @kazukim",
      "votes": null
    },
    {
      "id": "1215975",
      "postDate": "02/24/2021 04:29:59",
      "content": "<p>Thank you for your comment.<br>\nI hope I can help you even a little.</p>",
      "rawMarkdown": "Thank you for your comment.\nI hope I can help you even a little.",
      "votes": null
    },
    {
      "id": "1215993",
      "postDate": "02/24/2021 04:32:54",
      "content": "<p>Thank you.<br>\nI referred to your notebook.<br>\n<a href=\"https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903\" target=\"_blank\">https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903</a></p>",
      "rawMarkdown": "Thank you.\nI referred to your notebook.\nhttps://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903",
      "votes": null
    },
    {
      "id": "1216168",
      "postDate": "02/24/2021 06:16:52",
      "content": "<p>Thanks for sharing. How long did it take to train one epoch on EfficientNetB5? I'm seeing a lot of cases saying that it takes pretty long time, wonder how it had been for you. </p>",
      "rawMarkdown": "Thanks for sharing. How long did it take to train one epoch on EfficientNetB5? I'm seeing a lot of cases saying that it takes pretty long time, wonder how it had been for you.",
      "votes": null
    },
    {
      "id": "1216313",
      "postDate": "02/24/2021 07:51:31",
      "content": "<p>Thank you for your comment.</p>\n<p>EfficientNetB5 takes about 30 minutes per epoch.<br>\nTherefore, the total of training is 25 hours (about 5 hours per fold).<br>\nI did it on AWS.</p>",
      "rawMarkdown": "Thank you for your comment.\n\nEfficientNetB5 takes about 30 minutes per epoch.\nTherefore, the total of training is 25 hours (about 5 hours per fold).\nI did it on AWS.",
      "votes": null
    },
    {
      "id": "1216334",
      "postDate": "02/24/2021 08:05:18",
      "content": "<p>I see… since you've used AWS, I'm assuming the GPU is better than the one's provided on Colab or Kaggle? Must been pretty tough setting the environment up for model training. Great job!</p>",
      "rawMarkdown": "I see... since you've used AWS, I'm assuming the GPU is better than the one's provided on Colab or Kaggle? Must been pretty tough setting the environment up for model training. Great job!",
      "votes": null
    },
    {
      "id": "1216928",
      "postDate": "02/24/2021 16:10:39",
      "content": "<p>Great job! Thank you for sharing your experience!</p>",
      "rawMarkdown": "Great job! Thank you for sharing your experience!",
      "votes": null
    },
    {
      "id": "1218435",
      "postDate": "02/25/2021 21:33:24",
      "content": "<p>Great congratulations. <br>\nThis solution shows that even with the simplest methods (Adam, CrossEntropyLoss) it was possible to achieve high private LB score.</p>",
      "rawMarkdown": "Great congratulations. \nThis solution shows that even with the simplest methods (Adam, CrossEntropyLoss) it was possible to achieve high private LB score.",
      "votes": null
    },
    {
      "id": "1224594",
      "postDate": "03/02/2021 22:05:49",
      "content": "<p>Nice work, good summary )</p>",
      "rawMarkdown": "Nice work, good summary )",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1215939,
      "author_name": "hihunjin",
      "author_url": "",
      "post_date": "02/24/2021 04:17:46",
      "content": "<p>This model doesn't have that much tricks. Just use CrossEntropyLoss And Neither rigid regularization/generalization technique like TTA, Augmentation, nor Additional Extra Datasets.</p>\n<p>Today I learn,  Simple is The Best.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1215975,
          "author_name": "kazukim",
          "author_url": "",
          "post_date": "02/24/2021 04:29:59",
          "content": "<p>Thank you for your comment.<br>\nI hope I can help you even a little.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1215970,
      "author_name": "piantic",
      "author_url": "",
      "post_date": "02/24/2021 04:28:45",
      "content": "<p>Nice job! Simple is always best. <a href=\"https://www.kaggle.com/kazukim\" target=\"_blank\">@kazukim</a> </p>",
      "votes": null,
      "replies": [
        {
          "id": 1215993,
          "author_name": "kazukim",
          "author_url": "",
          "post_date": "02/24/2021 04:32:54",
          "content": "<p>Thank you.<br>\nI referred to your notebook.<br>\n<a href=\"https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903\" target=\"_blank\">https://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1216168,
      "author_name": "jiny333",
      "author_url": "",
      "post_date": "02/24/2021 06:16:52",
      "content": "<p>Thanks for sharing. How long did it take to train one epoch on EfficientNetB5? I'm seeing a lot of cases saying that it takes pretty long time, wonder how it had been for you. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1216313,
          "author_name": "kazukim",
          "author_url": "",
          "post_date": "02/24/2021 07:51:31",
          "content": "<p>Thank you for your comment.</p>\n<p>EfficientNetB5 takes about 30 minutes per epoch.<br>\nTherefore, the total of training is 25 hours (about 5 hours per fold).<br>\nI did it on AWS.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1216334,
          "author_name": "jiny333",
          "author_url": "",
          "post_date": "02/24/2021 08:05:18",
          "content": "<p>I see… since you've used AWS, I'm assuming the GPU is better than the one's provided on Colab or Kaggle? Must been pretty tough setting the environment up for model training. Great job!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1216928,
      "author_name": "omarchevska",
      "author_url": "",
      "post_date": "02/24/2021 16:10:39",
      "content": "<p>Great job! Thank you for sharing your experience!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1218435,
      "author_name": "aliabdin1",
      "author_url": "",
      "post_date": "02/25/2021 21:33:24",
      "content": "<p>Great congratulations. <br>\nThis solution shows that even with the simplest methods (Adam, CrossEntropyLoss) it was possible to achieve high private LB score.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1224594,
      "author_name": "vraizen",
      "author_url": "",
      "post_date": "03/02/2021 22:05:49",
      "content": "<p>Nice work, good summary )</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1215898": "## Acknowledgements\n   Thanks to Kaggle and hosts for holding this competition.\n   Thanks Kaggler!! This is my first silver medal :) \n\n## Result\n   - 30th\n   - Private Score 0.9010\n   - Public Score 0.9018\n\n## Summary\n    My final submission is ensemble of EfficientNetB5 with Noisy Student and ResNext50(32x4d).\n\n## Datasets\n    This competition datasets only. (I don't use 2019 datasets.)\n\n## Preprocessing\n    Image size 512\n    Augumentations\n     - ResizeCrop\n     - Horizontal,VerticalFlip\n     - OneOf(RandomBrightness, RandomContrast)\n     - OneOf (RandomBlur,MedianBlur,GaussianBlur)\n     - ShiftScaleRotate\n     - Normalize\n\n## Training\n    Optimizer Adam\n    Loss function CrossEntropyLoss\n\n## Base Model CV\n    CV is StratifiedKFold 5 folds.\n- EfficientNetB5\n    - CV average 0.8898\n    - Public 0.894\n    - Private 0.897\n- ResNext50(32x4d)\n    - CV average 0.8898\n    - Public 0.896\n    - Private 0.894\n\nhttps://github.com/kazukim10/kaggle_CasavaLeafDeseaseClassfication",
    "1215939": "This model doesn't have that much tricks. Just use CrossEntropyLoss And Neither rigid regularization/generalization technique like TTA, Augmentation, nor Additional Extra Datasets.\n\nToday I learn,  Simple is The Best.",
    "1215970": "Nice job! Simple is always best. @kazukim",
    "1215975": "Thank you for your comment.\nI hope I can help you even a little.",
    "1215993": "Thank you.\nI referred to your notebook.\nhttps://www.kaggle.com/piantic/no-tta-cassava-resnext50-32x4d-inference-lb0-903",
    "1216168": "Thanks for sharing. How long did it take to train one epoch on EfficientNetB5? I'm seeing a lot of cases saying that it takes pretty long time, wonder how it had been for you.",
    "1216313": "Thank you for your comment.\n\nEfficientNetB5 takes about 30 minutes per epoch.\nTherefore, the total of training is 25 hours (about 5 hours per fold).\nI did it on AWS.",
    "1216334": "I see... since you've used AWS, I'm assuming the GPU is better than the one's provided on Colab or Kaggle? Must been pretty tough setting the environment up for model training. Great job!",
    "1216928": "Great job! Thank you for sharing your experience!",
    "1218435": "Great congratulations. \nThis solution shows that even with the simplest methods (Adam, CrossEntropyLoss) it was possible to achieve high private LB score.",
    "1224594": "Nice work, good summary )"
  },
  "source": "meta"
}