{
  "id": 220593,
  "title": "Public 13th / Private 184th Solution",
  "url": "/competitions/cassava-leaf-disease-classification/writeups/master-breaker-public-13th-private-184th-solution",
  "author_name": "",
  "post_date": "2021-02-19T05:27:49.327Z",
  "votes": 15,
  "comment_count": 6,
  "views": 0,
  "content": "<h3>1. Overview</h3>\n<p>We selected two versions by increasing the variety of models and models with the highest score on the leaderboard.</p>\n<p><strong>Submission1 (Variety Models)</strong></p>\n<ul>\n<li>Public LB : 904 Private LB : 899</li>\n<li>Inference Code : <a href=\"https://www.kaggle.com/hihunjin/avengers-assemble-3-model10-tta2?scriptVersionId=54655767\" target=\"_blank\">https://www.kaggle.com/hihunjin/avengers-assemble-3-model10-tta2?scriptVersionId=54655767</a></li>\n</ul>\n<p><img src=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission1.png?raw=true\" alt=\"\"></p>\n<p><strong>Submission2 (Highest LB)</strong> </p>\n<ul>\n<li>Public LB : 908 Private LB : 897</li>\n<li>Inference Code : <a href=\"https://www.kaggle.com/chocozzz/sub4-effx3-regx3-tta-3?scriptVersionId=54655742\" target=\"_blank\">https://www.kaggle.com/chocozzz/sub4-effx3-regx3-tta-3?scriptVersionId=54655742</a></li>\n</ul>\n<p><img src=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission2.png?raw=true\" alt=\"Submission2\"></p>\n<h3>2. Tricks or Magics</h3>\n<ul>\n<li>Augmentations : LB 899 to 900 with GridMask</li>\n</ul>\n<pre><code>  def get_train_transforms():\n      return Compose([\n              RandomResizedCrop(CFG['img_size'], CFG['img_size']),\n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              ShiftScaleRotate(p=0.5),\n              HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),\n              RandomBrightnessContrast(brightness_limit=(-0.1,0.1), contrast_limit=(-0.1, 0.1), p=0.5),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              CoarseDropout(p=0.5),\n              GridMask(num_grid=3, p=0.5), # 899 to 900 \n              ToTensorV2(p=1.0),\n          ], p=1.)\n</code></pre>\n<ul>\n<li><p>Label Smoothing Loss : LB 900 to 901 </p>\n<ul>\n<li>alpha : 0.2 (When testing values between 0.05 and 0.3, 0.2 was the best score.)</li></ul></li>\n<li><p>SWA : LB 901 to 902 </p></li>\n<li><p>TTA Inference : LB 902 to 905 </p></li>\n</ul>\n<pre><code>  def get_inference_transforms():\n      return Compose([\n              OneOf([\n                  Resize(CFG['img_size'], CFG['img_size'], p=1.),\n                  CenterCrop(CFG['img_size'], CFG['img_size'], p=1.),\n                  RandomResizedCrop(CFG['img_size'], CFG['img_size'], p=1.)\n              ], p=1.), \n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              Resize(CFG['img_size'], CFG['img_size']),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              ToTensorV2(p=1.0),\n          ], p=1.)\n</code></pre>\n<ul>\n<li>only CenterCrop : 902 </li>\n<li>using RandomResized + CenterCrop : 904</li>\n<li>using Resize + CenterCrop + RandomResizedCrop : 905 </li>\n<li>Optimizer : Adam to Adamp (EfficientNet has reduced performance)<ul>\n<li>VIT : 901 to 902 </li>\n<li>RegNetY40 : 904 to 905 </li>\n<li>Eff : 905 to 904</li></ul></li>\n</ul>\n<p>Ensemble (Eff 905 / RegNetY 905 / VIT 902)</p>\n<ul>\n<li>EFF(2019+2020) + EFF(2020) + EFF(SEED 720) + RegNetY(2019+2020) + RegNetY(2020) + RegNetY(Distillation) : 908</li>\n<li>EFF(2019+2020) + EFF(2020) + RegNetY(2019+2020) + RegNetY(2020) : 907</li>\n<li>EFF + RegNetY : 907 </li>\n<li>EFF + VIT : 906 </li>\n<li>EFF + VIT + RegNetY : 906 </li>\n</ul>\n<h3>3. Trial and Error</h3>\n<ul>\n<li>Preprocessing : Noise Label Delete (using oof score), Clean Lab, Merging 2019 Dataset </li>\n<li>Augmentation : Cutmix, Cutout, Fmix, , 5 Crop etc.</li>\n<li>Loss : Bi-Tempered Logistic Loss, Focal Loss, Taylor Loss etc.</li>\n<li>Models : NFNet, BotNet, DeiT, Distillation, Fixmatch </li>\n<li>Optimizer : Adamw, Adamp (Sometimes it works for each model, and sometimes it doesn't) , RAdam, SAM </li>\n<li>Ensemble : SEED Ensemble, VIT + RegNetY40 + EfficientNet etc. </li>\n<li>Others : Batch Normalization Freeze, Hyper parameter Tuning (lr, gradient accum etc), Global Average Pooling to Concatenate Max Pooling and GAP etc. </li>\n</ul>\n<p>You can see our All Training code in the GitHub below. </p>\n<ul>\n<li>GitHub : <a href=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification\" target=\"_blank\">https://github.com/choco9966/Cassava-Leaf-Disease-Classification</a></li>\n</ul>\n<p>Lastly, I would like to say thank you to the team members who worked hard for 2 months.@Minyong Shin <a href=\"https://www.kaggle.com/bellwood\" target=\"_blank\">@bellwood</a> <a href=\"https://www.kaggle.com/hihunjin\" target=\"_blank\">@hihunjin</a> <a href=\"https://www.kaggle.com/cysohn\" target=\"_blank\">@cysohn</a> </p>",
  "messages": [
    {
      "id": "1209534",
      "postDate": "02/19/2021 00:17:55",
      "content": "<h3>1. Overview</h3>\n<p>We selected two versions by increasing the variety of models and models with the highest score on the leaderboard.</p>\n<p><strong>Submission1 (Variety Models)</strong></p>\n<ul>\n<li>Public LB : 904 Private LB : 899</li>\n<li>Inference Code : <a href=\"https://www.kaggle.com/hihunjin/avengers-assemble-3-model10-tta2?scriptVersionId=54655767\" target=\"_blank\">https://www.kaggle.com/hihunjin/avengers-assemble-3-model10-tta2?scriptVersionId=54655767</a></li>\n</ul>\n<p><img src=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission1.png?raw=true\" alt=\"\"></p>\n<p><strong>Submission2 (Highest LB)</strong> </p>\n<ul>\n<li>Public LB : 908 Private LB : 897</li>\n<li>Inference Code : <a href=\"https://www.kaggle.com/chocozzz/sub4-effx3-regx3-tta-3?scriptVersionId=54655742\" target=\"_blank\">https://www.kaggle.com/chocozzz/sub4-effx3-regx3-tta-3?scriptVersionId=54655742</a></li>\n</ul>\n<p><img src=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission2.png?raw=true\" alt=\"Submission2\"></p>\n<h3>2. Tricks or Magics</h3>\n<ul>\n<li>Augmentations : LB 899 to 900 with GridMask</li>\n</ul>\n<pre><code>  def get_train_transforms():\n      return Compose([\n              RandomResizedCrop(CFG['img_size'], CFG['img_size']),\n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              ShiftScaleRotate(p=0.5),\n              HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),\n              RandomBrightnessContrast(brightness_limit=(-0.1,0.1), contrast_limit=(-0.1, 0.1), p=0.5),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              CoarseDropout(p=0.5),\n              GridMask(num_grid=3, p=0.5), # 899 to 900 \n              ToTensorV2(p=1.0),\n          ], p=1.)\n</code></pre>\n<ul>\n<li><p>Label Smoothing Loss : LB 900 to 901 </p>\n<ul>\n<li>alpha : 0.2 (When testing values between 0.05 and 0.3, 0.2 was the best score.)</li></ul></li>\n<li><p>SWA : LB 901 to 902 </p></li>\n<li><p>TTA Inference : LB 902 to 905 </p></li>\n</ul>\n<pre><code>  def get_inference_transforms():\n      return Compose([\n              OneOf([\n                  Resize(CFG['img_size'], CFG['img_size'], p=1.),\n                  CenterCrop(CFG['img_size'], CFG['img_size'], p=1.),\n                  RandomResizedCrop(CFG['img_size'], CFG['img_size'], p=1.)\n              ], p=1.), \n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              Resize(CFG['img_size'], CFG['img_size']),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              ToTensorV2(p=1.0),\n          ], p=1.)\n</code></pre>\n<ul>\n<li>only CenterCrop : 902 </li>\n<li>using RandomResized + CenterCrop : 904</li>\n<li>using Resize + CenterCrop + RandomResizedCrop : 905 </li>\n<li>Optimizer : Adam to Adamp (EfficientNet has reduced performance)<ul>\n<li>VIT : 901 to 902 </li>\n<li>RegNetY40 : 904 to 905 </li>\n<li>Eff : 905 to 904</li></ul></li>\n</ul>\n<p>Ensemble (Eff 905 / RegNetY 905 / VIT 902)</p>\n<ul>\n<li>EFF(2019+2020) + EFF(2020) + EFF(SEED 720) + RegNetY(2019+2020) + RegNetY(2020) + RegNetY(Distillation) : 908</li>\n<li>EFF(2019+2020) + EFF(2020) + RegNetY(2019+2020) + RegNetY(2020) : 907</li>\n<li>EFF + RegNetY : 907 </li>\n<li>EFF + VIT : 906 </li>\n<li>EFF + VIT + RegNetY : 906 </li>\n</ul>\n<h3>3. Trial and Error</h3>\n<ul>\n<li>Preprocessing : Noise Label Delete (using oof score), Clean Lab, Merging 2019 Dataset </li>\n<li>Augmentation : Cutmix, Cutout, Fmix, , 5 Crop etc.</li>\n<li>Loss : Bi-Tempered Logistic Loss, Focal Loss, Taylor Loss etc.</li>\n<li>Models : NFNet, BotNet, DeiT, Distillation, Fixmatch </li>\n<li>Optimizer : Adamw, Adamp (Sometimes it works for each model, and sometimes it doesn't) , RAdam, SAM </li>\n<li>Ensemble : SEED Ensemble, VIT + RegNetY40 + EfficientNet etc. </li>\n<li>Others : Batch Normalization Freeze, Hyper parameter Tuning (lr, gradient accum etc), Global Average Pooling to Concatenate Max Pooling and GAP etc. </li>\n</ul>\n<p>You can see our All Training code in the GitHub below. </p>\n<ul>\n<li>GitHub : <a href=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification\" target=\"_blank\">https://github.com/choco9966/Cassava-Leaf-Disease-Classification</a></li>\n</ul>\n<p>Lastly, I would like to say thank you to the team members who worked hard for 2 months.@Minyong Shin <a href=\"https://www.kaggle.com/bellwood\" target=\"_blank\">@bellwood</a> <a href=\"https://www.kaggle.com/hihunjin\" target=\"_blank\">@hihunjin</a> <a href=\"https://www.kaggle.com/cysohn\" target=\"_blank\">@cysohn</a> </p>",
      "rawMarkdown": "### 1. Overview \n\nWe selected two versions by increasing the variety of models and models with the highest score on the leaderboard.\n\n**Submission1 (Variety Models)**\n\n- Public LB : 904 Private LB : 899\n- Inference Code : https://www.kaggle.com/hihunjin/avengers-assemble-3-model10-tta2?scriptVersionId=54655767\n\n![](https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission1.png?raw=true)\n\n**Submission2 (Highest LB)** \n\n- Public LB : 908 Private LB : 897\n- Inference Code : https://www.kaggle.com/chocozzz/sub4-effx3-regx3-tta-3?scriptVersionId=54655742\n\n![Submission2](https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission2.png?raw=true)\n\n### 2. Tricks or Magics \n\n- Augmentations : LB 899 to 900 with GridMask\n\n  ```python\n  def get_train_transforms():\n      return Compose([\n              RandomResizedCrop(CFG['img_size'], CFG['img_size']),\n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              ShiftScaleRotate(p=0.5),\n              HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),\n              RandomBrightnessContrast(brightness_limit=(-0.1,0.1), contrast_limit=(-0.1, 0.1), p=0.5),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              CoarseDropout(p=0.5),\n              GridMask(num_grid=3, p=0.5), # 899 to 900 \n              ToTensorV2(p=1.0),\n          ], p=1.)\n  ```\n\n- Label Smoothing Loss : LB 900 to 901 \n\n  - alpha : 0.2 (When testing values between 0.05 and 0.3, 0.2 was the best score.)\n\n- SWA : LB 901 to 902 \n\n- TTA Inference : LB 902 to 905 \n\n  ```python\n  def get_inference_transforms():\n      return Compose([\n              OneOf([\n                  Resize(CFG['img_size'], CFG['img_size'], p=1.),\n                  CenterCrop(CFG['img_size'], CFG['img_size'], p=1.),\n                  RandomResizedCrop(CFG['img_size'], CFG['img_size'], p=1.)\n              ], p=1.), \n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              Resize(CFG['img_size'], CFG['img_size']),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              ToTensorV2(p=1.0),\n          ], p=1.)\n  ```\n\n  - only CenterCrop : 902 \n  - using RandomResized + CenterCrop : 904\n  - using Resize + CenterCrop + RandomResizedCrop : 905 \n  - Optimizer : Adam to Adamp (EfficientNet has reduced performance)\n      - VIT : 901 to 902 \n      - RegNetY40 : 904 to 905 \n      - Eff : 905 to 904\n\n\nEnsemble (Eff 905 / RegNetY 905 / VIT 902)\n - EFF(2019+2020) + EFF(2020) + EFF(SEED 720) + RegNetY(2019+2020) + RegNetY(2020) + RegNetY(Distillation) : 908\n - EFF(2019+2020) + EFF(2020) + RegNetY(2019+2020) + RegNetY(2020) : 907\n - EFF + RegNetY : 907 \n - EFF + VIT : 906 \n - EFF + VIT + RegNetY : 906 \n\n### 3. Trial and Error \n\n- Preprocessing : Noise Label Delete (using oof score), Clean Lab, Merging 2019 Dataset \n- Augmentation : Cutmix, Cutout, Fmix, , 5 Crop etc.\n- Loss : Bi-Tempered Logistic Loss, Focal Loss, Taylor Loss etc.\n- Models : NFNet, BotNet, DeiT, Distillation, Fixmatch \n- Optimizer : Adamw, Adamp (Sometimes it works for each model, and sometimes it doesn't) , RAdam, SAM \n- Ensemble : SEED Ensemble, VIT + RegNetY40 + EfficientNet etc. \n- Others : Batch Normalization Freeze, Hyper parameter Tuning (lr, gradient accum etc), Global Average Pooling to Concatenate Max Pooling and GAP etc. \n\nYou can see our All Training code in the GitHub below. \n\n- GitHub : https://github.com/choco9966/Cassava-Leaf-Disease-Classification\n\nLastly, I would like to say thank you to the team members who worked hard for 2 months.@Minyong Shin @bellwood @hihunjin @cysohn",
      "votes": null
    },
    {
      "id": "1209576",
      "postDate": "02/19/2021 00:38:33",
      "content": "<p>고생하셨습니다~~ 다음에는 같이 한 대회 하시죠</p>",
      "rawMarkdown": "고생하셨습니다~~ 다음에는 같이 한 대회 하시죠",
      "votes": null
    },
    {
      "id": "1209588",
      "postDate": "02/19/2021 00:44:25",
      "content": "<p>통계왕형님도 메달하나가 떨어졌군요 ㅠㅠ 고생하셨어요~ 나중에 함 시간잡아서 얘기 나눠요 ㅋㅋㅋ</p>",
      "rawMarkdown": "통계왕형님도 메달하나가 떨어졌군요 ㅠㅠ 고생하셨어요~ 나중에 함 시간잡아서 얘기 나눠요 ㅋㅋㅋ",
      "votes": null
    },
    {
      "id": "1209641",
      "postDate": "02/19/2021 01:41:56",
      "content": "<p>Thanks for sharing great solution 👍<br>\n수고하셨습니다~!</p>",
      "rawMarkdown": "Thanks for sharing great solution 👍\n수고하셨습니다~!",
      "votes": null
    },
    {
      "id": "1209802",
      "postDate": "02/19/2021 03:43:52",
      "content": "<p>고생하셨습니다~~!!👍👍</p>",
      "rawMarkdown": "고생하셨습니다~~!!👍👍",
      "votes": null
    },
    {
      "id": "1209908",
      "postDate": "02/19/2021 05:24:40",
      "content": "<p>현우님 링크드인 포스트 잘 봤습니다! 최종 제출에서 순위가 많이 하락했군요 ㅠㅠ<br>\n고생 많으셨습니다~<br>\n(Discussion에 올려주신 이미지가 저한테는 보이지 않네요…!)</p>",
      "rawMarkdown": "현우님 링크드인 포스트 잘 봤습니다! 최종 제출에서 순위가 많이 하락했군요 ㅠㅠ\n고생 많으셨습니다~\n(Discussion에 올려주신 이미지가 저한테는 보이지 않네요...!)",
      "votes": null
    },
    {
      "id": "1209914",
      "postDate": "02/19/2021 05:29:26",
      "content": "<p>헉 감사합니다. 최종에서 많이 떨어져가지고 ㄲㄲ… 파일 수정했는데 아직도 사진 안보이시나요?? <a href=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification/tree/main/image\" target=\"_blank\">https://github.com/choco9966/Cassava-Leaf-Disease-Classification/tree/main/image</a> 요 링크에 있는 submission 사진 두개였는데 캐글 지금 이미지 업로드가 이상한 것 같네요 😭😭</p>",
      "rawMarkdown": "헉 감사합니다. 최종에서 많이 떨어져가지고 ㄲㄲ... 파일 수정했는데 아직도 사진 안보이시나요?? https://github.com/choco9966/Cassava-Leaf-Disease-Classification/tree/main/image 요 링크에 있는 submission 사진 두개였는데 캐글 지금 이미지 업로드가 이상한 것 같네요 😭😭",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1209576,
      "author_name": "deepkim",
      "author_url": "",
      "post_date": "02/19/2021 00:38:33",
      "content": "<p>고생하셨습니다~~ 다음에는 같이 한 대회 하시죠</p>",
      "votes": null,
      "replies": [
        {
          "id": 1209588,
          "author_name": "chocozzz",
          "author_url": "",
          "post_date": "02/19/2021 00:44:25",
          "content": "<p>통계왕형님도 메달하나가 떨어졌군요 ㅠㅠ 고생하셨어요~ 나중에 함 시간잡아서 얘기 나눠요 ㅋㅋㅋ</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1209641,
      "author_name": "nevret93",
      "author_url": "",
      "post_date": "02/19/2021 01:41:56",
      "content": "<p>Thanks for sharing great solution 👍<br>\n수고하셨습니다~!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1209802,
          "author_name": "chocozzz",
          "author_url": "",
          "post_date": "02/19/2021 03:43:52",
          "content": "<p>고생하셨습니다~~!!👍👍</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1209908,
      "author_name": "leeyj0511",
      "author_url": "",
      "post_date": "02/19/2021 05:24:40",
      "content": "<p>현우님 링크드인 포스트 잘 봤습니다! 최종 제출에서 순위가 많이 하락했군요 ㅠㅠ<br>\n고생 많으셨습니다~<br>\n(Discussion에 올려주신 이미지가 저한테는 보이지 않네요…!)</p>",
      "votes": null,
      "replies": [
        {
          "id": 1209914,
          "author_name": "chocozzz",
          "author_url": "",
          "post_date": "02/19/2021 05:29:26",
          "content": "<p>헉 감사합니다. 최종에서 많이 떨어져가지고 ㄲㄲ… 파일 수정했는데 아직도 사진 안보이시나요?? <a href=\"https://github.com/choco9966/Cassava-Leaf-Disease-Classification/tree/main/image\" target=\"_blank\">https://github.com/choco9966/Cassava-Leaf-Disease-Classification/tree/main/image</a> 요 링크에 있는 submission 사진 두개였는데 캐글 지금 이미지 업로드가 이상한 것 같네요 😭😭</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1209534": "### 1. Overview \n\nWe selected two versions by increasing the variety of models and models with the highest score on the leaderboard.\n\n**Submission1 (Variety Models)**\n\n- Public LB : 904 Private LB : 899\n- Inference Code : https://www.kaggle.com/hihunjin/avengers-assemble-3-model10-tta2?scriptVersionId=54655767\n\n![](https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission1.png?raw=true)\n\n**Submission2 (Highest LB)** \n\n- Public LB : 908 Private LB : 897\n- Inference Code : https://www.kaggle.com/chocozzz/sub4-effx3-regx3-tta-3?scriptVersionId=54655742\n\n![Submission2](https://github.com/choco9966/Cassava-Leaf-Disease-Classification/blob/main/image/submission2.png?raw=true)\n\n### 2. Tricks or Magics \n\n- Augmentations : LB 899 to 900 with GridMask\n\n  ```python\n  def get_train_transforms():\n      return Compose([\n              RandomResizedCrop(CFG['img_size'], CFG['img_size']),\n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              ShiftScaleRotate(p=0.5),\n              HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),\n              RandomBrightnessContrast(brightness_limit=(-0.1,0.1), contrast_limit=(-0.1, 0.1), p=0.5),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              CoarseDropout(p=0.5),\n              GridMask(num_grid=3, p=0.5), # 899 to 900 \n              ToTensorV2(p=1.0),\n          ], p=1.)\n  ```\n\n- Label Smoothing Loss : LB 900 to 901 \n\n  - alpha : 0.2 (When testing values between 0.05 and 0.3, 0.2 was the best score.)\n\n- SWA : LB 901 to 902 \n\n- TTA Inference : LB 902 to 905 \n\n  ```python\n  def get_inference_transforms():\n      return Compose([\n              OneOf([\n                  Resize(CFG['img_size'], CFG['img_size'], p=1.),\n                  CenterCrop(CFG['img_size'], CFG['img_size'], p=1.),\n                  RandomResizedCrop(CFG['img_size'], CFG['img_size'], p=1.)\n              ], p=1.), \n              Transpose(p=0.5),\n              HorizontalFlip(p=0.5),\n              VerticalFlip(p=0.5),\n              Resize(CFG['img_size'], CFG['img_size']),\n              Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),\n              ToTensorV2(p=1.0),\n          ], p=1.)\n  ```\n\n  - only CenterCrop : 902 \n  - using RandomResized + CenterCrop : 904\n  - using Resize + CenterCrop + RandomResizedCrop : 905 \n  - Optimizer : Adam to Adamp (EfficientNet has reduced performance)\n      - VIT : 901 to 902 \n      - RegNetY40 : 904 to 905 \n      - Eff : 905 to 904\n\n\nEnsemble (Eff 905 / RegNetY 905 / VIT 902)\n - EFF(2019+2020) + EFF(2020) + EFF(SEED 720) + RegNetY(2019+2020) + RegNetY(2020) + RegNetY(Distillation) : 908\n - EFF(2019+2020) + EFF(2020) + RegNetY(2019+2020) + RegNetY(2020) : 907\n - EFF + RegNetY : 907 \n - EFF + VIT : 906 \n - EFF + VIT + RegNetY : 906 \n\n### 3. Trial and Error \n\n- Preprocessing : Noise Label Delete (using oof score), Clean Lab, Merging 2019 Dataset \n- Augmentation : Cutmix, Cutout, Fmix, , 5 Crop etc.\n- Loss : Bi-Tempered Logistic Loss, Focal Loss, Taylor Loss etc.\n- Models : NFNet, BotNet, DeiT, Distillation, Fixmatch \n- Optimizer : Adamw, Adamp (Sometimes it works for each model, and sometimes it doesn't) , RAdam, SAM \n- Ensemble : SEED Ensemble, VIT + RegNetY40 + EfficientNet etc. \n- Others : Batch Normalization Freeze, Hyper parameter Tuning (lr, gradient accum etc), Global Average Pooling to Concatenate Max Pooling and GAP etc. \n\nYou can see our All Training code in the GitHub below. \n\n- GitHub : https://github.com/choco9966/Cassava-Leaf-Disease-Classification\n\nLastly, I would like to say thank you to the team members who worked hard for 2 months.@Minyong Shin @bellwood @hihunjin @cysohn",
    "1209576": "고생하셨습니다~~ 다음에는 같이 한 대회 하시죠",
    "1209588": "통계왕형님도 메달하나가 떨어졌군요 ㅠㅠ 고생하셨어요~ 나중에 함 시간잡아서 얘기 나눠요 ㅋㅋㅋ",
    "1209641": "Thanks for sharing great solution 👍\n수고하셨습니다~!",
    "1209802": "고생하셨습니다~~!!👍👍",
    "1209908": "현우님 링크드인 포스트 잘 봤습니다! 최종 제출에서 순위가 많이 하락했군요 ㅠㅠ\n고생 많으셨습니다~\n(Discussion에 올려주신 이미지가 저한테는 보이지 않네요...!)",
    "1209914": "헉 감사합니다. 최종에서 많이 떨어져가지고 ㄲㄲ... 파일 수정했는데 아직도 사진 안보이시나요?? https://github.com/choco9966/Cassava-Leaf-Disease-Classification/tree/main/image 요 링크에 있는 submission 사진 두개였는데 캐글 지금 이미지 업로드가 이상한 것 같네요 😭😭"
  },
  "source": "meta"
}