{
  "id": 220648,
  "title": "#28th place solution",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/220648",
  "author_name": "",
  "post_date": "2021-02-19T04:38:41.403178700Z",
  "votes": 10,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi Kagglers,<br>\nFirst of all, congratulations to all the winners and did a great job on all participants' hard works.<br>\nThis is the first time my team has joined a Computer Vision competition and with little luck we had a sliver medal and the most important things, we learned so much from this competition. <br>\nHere, we share our solution that got #28th place.</p>\n<ol>\n<li>Data</li>\n</ol>\n<ul>\n<li>2019 + 2020 dataset </li>\n<li>Remove duplicate:<ul>\n<li>Vector embedding from B4-training on imagenet</li>\n<li>Using DBSCAN (eps=3) to remove duplicate=&gt; total 29409 image)</li></ul></li>\n<li>Cleaning noisy data:<ul>\n<li>Training based model Efficientnet B4 on 2019+2020 dataset</li>\n<li>Predict on 2020 dataset -&gt; hand-label image with truth / predict = 0 or 4 and wrong predict</li></ul></li>\n</ul>\n<ol>\n<li>Model</li>\n</ol>\n<ul>\n<li>Efficient B4</li>\n<li>ViT based 16*384</li>\n<li>Inception</li>\n<li>Xception</li>\n</ul>\n<ol>\n<li>Training strategy</li>\n</ol>\n<ul>\n<li>Augmentations:<ul>\n<li>Light augmentation, no cutmix, fmix…</li></ul></li>\n<li>Loss function:<ul>\n<li>Taylor Crossentropy + labelsmoothing</li>\n<li>Customize Focal loss + labelsmoothing</li>\n<li>Bi-tempred loss + labelsmoothing </li></ul></li>\n<li><a href=\"https://arxiv.org/abs/1906.06423\" target=\"_blank\">Fix resolution</a>: <ul>\n<li>training with 384 *384 image</li>\n<li>fine tuning (training last batch norm and classifier layer) with 512*512<br>\n=&gt; faster training and better model for both CV and LB</li></ul></li>\n<li>5 fold, 2020 for train and valid, 2019 for train only</li>\n<li>Cosineannealing warm start + adam</li>\n</ul>\n<ol>\n<li>Inference</li>\n</ol>\n<ul>\n<li>5x tta RandomResizedCrop(scale=0.3, 1)</li>\n<li>Average probability ViT, Inception, Xception, EfficientnetB4</li>\n</ul>\n<ol>\n<li>Result</li>\n</ol>\n<ul>\n<li>Single model B4 (single fold, notta): 0.901 public LB, 899 private LB</li>\n<li>Ensemble (ViT, Inception, Xception, EffB4) average probability with 5xtta: 90.2 public LB,  90.1 private LB<br>\n=&gt; Ensemble and large scale crop get stable score</li>\n</ul>\n<p>From <strong>Train4Ever</strong> team - <strong>Viettel Governance Department</strong><br>\n<a href=\"https://www.kaggle.com/tungvs\" target=\"_blank\">@tungvs</a> <a href=\"https://www.kaggle.com/hoangtubk\" target=\"_blank\">@hoangtubk</a> <a href=\"https://www.kaggle.com/namgalielei\" target=\"_blank\">@namgalielei</a> <a href=\"https://www.kaggle.com/tuyendam\" target=\"_blank\">@tuyendam</a> </p>",
  "messages": [
    {
      "id": "1209863",
      "postDate": "02/19/2021 04:38:41",
      "content": "<p>Hi Kagglers,<br>\nFirst of all, congratulations to all the winners and did a great job on all participants' hard works.<br>\nThis is the first time my team has joined a Computer Vision competition and with little luck we had a sliver medal and the most important things, we learned so much from this competition. <br>\nHere, we share our solution that got #28th place.</p>\n<ol>\n<li>Data</li>\n</ol>\n<ul>\n<li>2019 + 2020 dataset </li>\n<li>Remove duplicate:<ul>\n<li>Vector embedding from B4-training on imagenet</li>\n<li>Using DBSCAN (eps=3) to remove duplicate=&gt; total 29409 image)</li></ul></li>\n<li>Cleaning noisy data:<ul>\n<li>Training based model Efficientnet B4 on 2019+2020 dataset</li>\n<li>Predict on 2020 dataset -&gt; hand-label image with truth / predict = 0 or 4 and wrong predict</li></ul></li>\n</ul>\n<ol>\n<li>Model</li>\n</ol>\n<ul>\n<li>Efficient B4</li>\n<li>ViT based 16*384</li>\n<li>Inception</li>\n<li>Xception</li>\n</ul>\n<ol>\n<li>Training strategy</li>\n</ol>\n<ul>\n<li>Augmentations:<ul>\n<li>Light augmentation, no cutmix, fmix…</li></ul></li>\n<li>Loss function:<ul>\n<li>Taylor Crossentropy + labelsmoothing</li>\n<li>Customize Focal loss + labelsmoothing</li>\n<li>Bi-tempred loss + labelsmoothing </li></ul></li>\n<li><a href=\"https://arxiv.org/abs/1906.06423\" target=\"_blank\">Fix resolution</a>: <ul>\n<li>training with 384 *384 image</li>\n<li>fine tuning (training last batch norm and classifier layer) with 512*512<br>\n=&gt; faster training and better model for both CV and LB</li></ul></li>\n<li>5 fold, 2020 for train and valid, 2019 for train only</li>\n<li>Cosineannealing warm start + adam</li>\n</ul>\n<ol>\n<li>Inference</li>\n</ol>\n<ul>\n<li>5x tta RandomResizedCrop(scale=0.3, 1)</li>\n<li>Average probability ViT, Inception, Xception, EfficientnetB4</li>\n</ul>\n<ol>\n<li>Result</li>\n</ol>\n<ul>\n<li>Single model B4 (single fold, notta): 0.901 public LB, 899 private LB</li>\n<li>Ensemble (ViT, Inception, Xception, EffB4) average probability with 5xtta: 90.2 public LB,  90.1 private LB<br>\n=&gt; Ensemble and large scale crop get stable score</li>\n</ul>\n<p>From <strong>Train4Ever</strong> team - <strong>Viettel Governance Department</strong><br>\n<a href=\"https://www.kaggle.com/tungvs\" target=\"_blank\">@tungvs</a> <a href=\"https://www.kaggle.com/hoangtubk\" target=\"_blank\">@hoangtubk</a> <a href=\"https://www.kaggle.com/namgalielei\" target=\"_blank\">@namgalielei</a> <a href=\"https://www.kaggle.com/tuyendam\" target=\"_blank\">@tuyendam</a> </p>",
      "rawMarkdown": "Hi Kagglers,\nFirst of all, congratulations to all the winners and did a great job on all participants' hard works.\nThis is the first time my team has joined a Computer Vision competition and with little luck we had a sliver medal and the most important things, we learned so much from this competition. \nHere, we share our solution that got #28th place.\n1.\tData\n-\t2019 + 2020 dataset \n-\tRemove duplicate:\n    - Vector embedding from B4-training on imagenet\n    - Using DBSCAN (eps=3) to remove duplicate=> total 29409 image)\n-\tCleaning noisy data:\n    - Training based model Efficientnet B4 on 2019+2020 dataset\n    - Predict on 2020 dataset -> hand-label image with truth / predict = 0 or 4 and wrong predict\n2.\tModel\n-\tEfficient B4\n-\tViT based 16*384\n-\tInception\n-\tXception\n3.\tTraining strategy\n-\tAugmentations:\n    - Light augmentation, no cutmix, fmix…\n-\tLoss function:\n    - Taylor Crossentropy + labelsmoothing\n    - Customize Focal loss + labelsmoothing\n    - Bi-tempred loss + labelsmoothing \n-     [Fix resolution](https://arxiv.org/abs/1906.06423): \n    - training with 384 *384 image\n    - fine tuning (training last batch norm and classifier layer) with 512*512\n    => faster training and better model for both CV and LB\n-\t5 fold, 2020 for train and valid, 2019 for train only\n-\tCosineannealing warm start + adam\n4.\tInference\n-\t5x tta RandomResizedCrop(scale=0.3, 1)\n-\tAverage probability ViT, Inception, Xception, EfficientnetB4\n5.\tResult\n-\tSingle model B4 (single fold, notta): 0.901 public LB, 899 private LB\n-\tEnsemble (ViT, Inception, Xception, EffB4) average probability with 5xtta: 90.2 public LB,  90.1 private LB\n => Ensemble and large scale crop get stable score\n\nFrom **Train4Ever** team - **Viettel Governance Department**\n@tungvs @hoangtubk @namgalielei @tuyendam",
      "votes": null
    },
    {
      "id": "1209905",
      "postDate": "02/19/2021 05:21:52",
      "content": "<p>Nice work！ Congratulation！</p>",
      "rawMarkdown": "Nice work！ Congratulation！",
      "votes": null
    },
    {
      "id": "1209906",
      "postDate": "02/19/2021 05:22:44",
      "content": "<p>Congrats on 28th and solo silver medal! </p>\n<p>Cleaning noisy data and finetuning with 512x512 are great ideas. <a href=\"https://www.kaggle.com/researchbntz\" target=\"_blank\">@researchbntz</a> </p>",
      "rawMarkdown": "Congrats on 28th and solo silver medal! \n\nCleaning noisy data and finetuning with 512x512 are great ideas. @researchbntz",
      "votes": null
    },
    {
      "id": "1209922",
      "postDate": "02/19/2021 05:35:13",
      "content": "<p>Thanks… I also learned alot from your notebook and  discussion in this competition.</p>",
      "rawMarkdown": "Thanks... I also learned alot from your notebook and  discussion in this competition.",
      "votes": null
    },
    {
      "id": "1209955",
      "postDate": "02/19/2021 06:00:46",
      "content": "<p>Congrats on strong finish <a href=\"https://www.kaggle.com/researchbntz\" target=\"_blank\">@researchbntz</a> and team</p>",
      "rawMarkdown": "Congrats on strong finish @researchbntz and team",
      "votes": null
    },
    {
      "id": "1211028",
      "postDate": "02/19/2021 22:45:30",
      "content": "<p>How come I can't see his team and his solution anymore?</p>",
      "rawMarkdown": "How come I can't see his team and his solution anymore?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1209905,
      "author_name": "ouyangouy",
      "author_url": "",
      "post_date": "02/19/2021 05:21:52",
      "content": "<p>Nice work！ Congratulation！</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1209906,
      "author_name": "piantic",
      "author_url": "",
      "post_date": "02/19/2021 05:22:44",
      "content": "<p>Congrats on 28th and solo silver medal! </p>\n<p>Cleaning noisy data and finetuning with 512x512 are great ideas. <a href=\"https://www.kaggle.com/researchbntz\" target=\"_blank\">@researchbntz</a> </p>",
      "votes": null,
      "replies": [
        {
          "id": 1209922,
          "author_name": "researchbntz",
          "author_url": "",
          "post_date": "02/19/2021 05:35:13",
          "content": "<p>Thanks… I also learned alot from your notebook and  discussion in this competition.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1209955,
      "author_name": "duykhanh99",
      "author_url": "",
      "post_date": "02/19/2021 06:00:46",
      "content": "<p>Congrats on strong finish <a href=\"https://www.kaggle.com/researchbntz\" target=\"_blank\">@researchbntz</a> and team</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1211028,
      "author_name": "andyjianzhou",
      "author_url": "",
      "post_date": "02/19/2021 22:45:30",
      "content": "<p>How come I can't see his team and his solution anymore?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1209863": "Hi Kagglers,\nFirst of all, congratulations to all the winners and did a great job on all participants' hard works.\nThis is the first time my team has joined a Computer Vision competition and with little luck we had a sliver medal and the most important things, we learned so much from this competition. \nHere, we share our solution that got #28th place.\n1.\tData\n-\t2019 + 2020 dataset \n-\tRemove duplicate:\n    - Vector embedding from B4-training on imagenet\n    - Using DBSCAN (eps=3) to remove duplicate=> total 29409 image)\n-\tCleaning noisy data:\n    - Training based model Efficientnet B4 on 2019+2020 dataset\n    - Predict on 2020 dataset -> hand-label image with truth / predict = 0 or 4 and wrong predict\n2.\tModel\n-\tEfficient B4\n-\tViT based 16*384\n-\tInception\n-\tXception\n3.\tTraining strategy\n-\tAugmentations:\n    - Light augmentation, no cutmix, fmix…\n-\tLoss function:\n    - Taylor Crossentropy + labelsmoothing\n    - Customize Focal loss + labelsmoothing\n    - Bi-tempred loss + labelsmoothing \n-     [Fix resolution](https://arxiv.org/abs/1906.06423): \n    - training with 384 *384 image\n    - fine tuning (training last batch norm and classifier layer) with 512*512\n    => faster training and better model for both CV and LB\n-\t5 fold, 2020 for train and valid, 2019 for train only\n-\tCosineannealing warm start + adam\n4.\tInference\n-\t5x tta RandomResizedCrop(scale=0.3, 1)\n-\tAverage probability ViT, Inception, Xception, EfficientnetB4\n5.\tResult\n-\tSingle model B4 (single fold, notta): 0.901 public LB, 899 private LB\n-\tEnsemble (ViT, Inception, Xception, EffB4) average probability with 5xtta: 90.2 public LB,  90.1 private LB\n => Ensemble and large scale crop get stable score\n\nFrom **Train4Ever** team - **Viettel Governance Department**\n@tungvs @hoangtubk @namgalielei @tuyendam",
    "1209905": "Nice work！ Congratulation！",
    "1209906": "Congrats on 28th and solo silver medal! \n\nCleaning noisy data and finetuning with 512x512 are great ideas. @researchbntz",
    "1209922": "Thanks... I also learned alot from your notebook and  discussion in this competition.",
    "1209955": "Congrats on strong finish @researchbntz and team",
    "1211028": "How come I can't see his team and his solution anymore?"
  },
  "source": "meta"
}