{
  "id": 220900,
  "title": "40th place solution(4 model ensemble)",
  "url": "/competitions/cassava-leaf-disease-classification/writeups/c-c-40th-place-solution-4-model-ensemble",
  "author_name": "",
  "post_date": "2021-02-20T05:22:13.823Z",
  "votes": 8,
  "comment_count": 1,
  "views": 0,
  "content": "<p>First of all,I'd like to say thanks to <a href=\"https://www.kaggle.com/qi0239\" target=\"_blank\">@qi0239</a> <a href=\"https://www.kaggle.com/bilogo\" target=\"_blank\">@bilogo</a>  <a href=\"https://www.kaggle.com/ageage\" target=\"_blank\">@ageage</a>  <a href=\"https://www.kaggle.com/hsdtlp\" target=\"_blank\">@hsdtlp</a> ,this is our first match in kaggle,they spend much time finding suitable method and persisted till the end.</p>\n<h1>score</h1>\n<ul>\n<li>Public Leaderboard   0.9050</li>\n<li>Private Leaderboard  0.9007</li>\n</ul>\n<h1>training details</h1>\n<ul>\n<li>data:  2020 data with 5fold</li>\n<li>Image_size:  512x512</li>\n<li>Models: efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50</li>\n<li>Loss function: bi-tempered loss with label smooth  </li>\n<li>Optimizer: AdamW</li>\n<li>Lr scheduler: Cosine Annealing with Warm Restarts</li>\n<li>Augmentations: HorizontalFlip, VerticalFlip,ShiftScaleRotate,mixup,snapmix(when we using HueSaturationValue and RandomBrightnessContrast,it improve our cv,but lower lb)</li>\n<li>TTA: ShiftScaleRotate,Flip</li>\n</ul>\n<h1>ensemble</h1>\n<p>we using 4 models :efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50.like other teams,single model didn't performance as good as ensemble.At first,we using efficientnet b3 &amp; se_resnext50,by this way,we get higest public lb 0.900.when finding ensemble working,we continue test efficientnet-b4 and add it in our ensemble,in this discussion<a href=\"url\" target=\"_blank\">https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/213183</a>,we notice repvgg model,and we tried testing type A0-B3,find type B2g4 have best performance in cv and public lb,by adding repvgg,it help our public lb +0.003.</p>\n<h1>methods that didn't work</h1>\n<ul>\n<li>remove noisy images,we using PHash to find same images and remove similar images,both cv and lb become lower</li>\n<li>Pseudo-Labelling</li>\n<li>concat avgpool feature &amp; maxpool feature</li>\n</ul>\n<h1>in this match,I have learnt a lot,people here are enthusiasm,their Reply in the discussion gave me a lot of help  in the game.Best wish!</h1>",
  "messages": [
    {
      "id": "1211159",
      "postDate": "02/20/2021 02:26:27",
      "content": "<p>First of all,I'd like to say thanks to <a href=\"https://www.kaggle.com/qi0239\" target=\"_blank\">@qi0239</a> <a href=\"https://www.kaggle.com/bilogo\" target=\"_blank\">@bilogo</a>  <a href=\"https://www.kaggle.com/ageage\" target=\"_blank\">@ageage</a>  <a href=\"https://www.kaggle.com/hsdtlp\" target=\"_blank\">@hsdtlp</a> ,this is our first match in kaggle,they spend much time finding suitable method and persisted till the end.</p>\n<h1>score</h1>\n<ul>\n<li>Public Leaderboard   0.9050</li>\n<li>Private Leaderboard  0.9007</li>\n</ul>\n<h1>training details</h1>\n<ul>\n<li>data:  2020 data with 5fold</li>\n<li>Image_size:  512x512</li>\n<li>Models: efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50</li>\n<li>Loss function: bi-tempered loss with label smooth  </li>\n<li>Optimizer: AdamW</li>\n<li>Lr scheduler: Cosine Annealing with Warm Restarts</li>\n<li>Augmentations: HorizontalFlip, VerticalFlip,ShiftScaleRotate,mixup,snapmix(when we using HueSaturationValue and RandomBrightnessContrast,it improve our cv,but lower lb)</li>\n<li>TTA: ShiftScaleRotate,Flip</li>\n</ul>\n<h1>ensemble</h1>\n<p>we using 4 models :efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50.like other teams,single model didn't performance as good as ensemble.At first,we using efficientnet b3 &amp; se_resnext50,by this way,we get higest public lb 0.900.when finding ensemble working,we continue test efficientnet-b4 and add it in our ensemble,in this discussion<a href=\"url\" target=\"_blank\">https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/213183</a>,we notice repvgg model,and we tried testing type A0-B3,find type B2g4 have best performance in cv and public lb,by adding repvgg,it help our public lb +0.003.</p>\n<h1>methods that didn't work</h1>\n<ul>\n<li>remove noisy images,we using PHash to find same images and remove similar images,both cv and lb become lower</li>\n<li>Pseudo-Labelling</li>\n<li>concat avgpool feature &amp; maxpool feature</li>\n</ul>\n<h1>in this match,I have learnt a lot,people here are enthusiasm,their Reply in the discussion gave me a lot of help  in the game.Best wish!</h1>",
      "rawMarkdown": "First of all,I'd like to say thanks to @qi0239 @bilogo  @ageage  @hsdtlp ,this is our first match in kaggle,they spend much time finding suitable method and persisted till the end.\n# score\n- Public Leaderboard   0.9050\n- Private Leaderboard  0.9007\n# training details\n- data:  2020 data with 5fold\n- Image_size:  512x512\n- Models: efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50\n\n- Loss function: bi-tempered loss with label smooth  \n\n- Optimizer: AdamW\n\n- Lr scheduler: Cosine Annealing with Warm Restarts\n- Augmentations: HorizontalFlip, VerticalFlip,ShiftScaleRotate,mixup,snapmix(when we using HueSaturationValue and RandomBrightnessContrast,it improve our cv,but lower lb)\n- TTA: ShiftScaleRotate,Flip\n# ensemble\nwe using 4 models :efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50.like other teams,single model didn't performance as good as ensemble.At first,we using efficientnet b3 & se_resnext50,by this way,we get higest public lb 0.900.when finding ensemble working,we continue test efficientnet-b4 and add it in our ensemble,in this discussion[https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/213183](url),we notice repvgg model,and we tried testing type A0-B3,find type B2g4 have best performance in cv and public lb,by adding repvgg,it help our public lb +0.003.\n\n# methods that didn't work\n- remove noisy images,we using PHash to find same images and remove similar images,both cv and lb become lower\n- Pseudo-Labelling\n- concat avgpool feature & maxpool feature\n\n\n# in this match,I have learnt a lot,people here are enthusiasm,their Reply in the discussion gave me a lot of help  in the game.Best wish!",
      "votes": null
    },
    {
      "id": "1211470",
      "postDate": "02/20/2021 09:07:47",
      "content": "<p>great work!</p>",
      "rawMarkdown": "great work!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1211470,
      "author_name": "wonjunpark",
      "author_url": "",
      "post_date": "02/20/2021 09:07:47",
      "content": "<p>great work!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1211159": "First of all,I'd like to say thanks to @qi0239 @bilogo  @ageage  @hsdtlp ,this is our first match in kaggle,they spend much time finding suitable method and persisted till the end.\n# score\n- Public Leaderboard   0.9050\n- Private Leaderboard  0.9007\n# training details\n- data:  2020 data with 5fold\n- Image_size:  512x512\n- Models: efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50\n\n- Loss function: bi-tempered loss with label smooth  \n\n- Optimizer: AdamW\n\n- Lr scheduler: Cosine Annealing with Warm Restarts\n- Augmentations: HorizontalFlip, VerticalFlip,ShiftScaleRotate,mixup,snapmix(when we using HueSaturationValue and RandomBrightnessContrast,it improve our cv,but lower lb)\n- TTA: ShiftScaleRotate,Flip\n# ensemble\nwe using 4 models :efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50.like other teams,single model didn't performance as good as ensemble.At first,we using efficientnet b3 & se_resnext50,by this way,we get higest public lb 0.900.when finding ensemble working,we continue test efficientnet-b4 and add it in our ensemble,in this discussion[https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/213183](url),we notice repvgg model,and we tried testing type A0-B3,find type B2g4 have best performance in cv and public lb,by adding repvgg,it help our public lb +0.003.\n\n# methods that didn't work\n- remove noisy images,we using PHash to find same images and remove similar images,both cv and lb become lower\n- Pseudo-Labelling\n- concat avgpool feature & maxpool feature\n\n\n# in this match,I have learnt a lot,people here are enthusiasm,their Reply in the discussion gave me a lot of help  in the game.Best wish!",
    "1211470": "great work!"
  },
  "source": "meta"
}