{
  "id": 355213,
  "title": "27th Simple Solution",
  "url": "/competitions/hubmap-organ-segmentation/discussion/355213",
  "author_name": "JCW2023",
  "post_date": "2022-09-26T01:56:49.302000",
  "votes": 9,
  "comment_count": 0,
  "views": 0,
  "content": "<p>First of all, thank the organizers for hosting this competition. <br>\nWe also thank <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> for his discussions and notebook.</p>\n<p><strong>Models</strong><br>\nCoat_small+daformer,size=768,folds=5,swa<br>\nCoat_small+daformer,size=768,folds=5,External data , swa <br>\nCoat_lite_medium+daformer,size=768,folds=5,swa<br>\nCoat_lite_medium+daformer,size=768,all data,swa<br>\nCoat_small(level5)+daformer_conv1*1,size=1152, folds=5, swa<br>\npvt_v2_b4+daformer_conv3x3,size=768,folds=5,swa<br>\npvt_v2_b4+daformer_conv3x3,size=768,all data,swa<br>\npvt_v2_b4+smp_unet,size=768,folds=5,swa<br>\npvt_v2_b4+smp_unet,size=768,all data,swa</p>\n<p><strong>Augmentations</strong><br>\nRandom flip, rotate, noise, HueSaturationValue, affine..</p>\n<p><strong>Data Processing</strong><br>\nWe have selected three methods of training data:<br>\nThe original data is divided into training data and validation data for training. <br>\nThe full data is trained. <br>\nThe external data is labeled and then added to the original data for training</p>\n<p><strong>Inference</strong><br>\nEnsemble<br>\nCoat_small+daformer:fold0 fold3<br>\nCoat_lite_medium+daformer: fold2 fold3 fold4 all_data<br>\nCoat_small(level5)+daformer_conv1*1 :fold0_1152<br>\npvt_v2_b4+daformer_conv3x3:fold1,fold2,fold3,fold4,all_data,<br>\nfold3_384,fold2_384,fold3_1024<br>\npvt_v2_b4+smp_unet:fold0,fold2,all_data</p>\n<p>TTA<br>\nFlip,Rorate</p>\n<p>Organ Threshold<br>\n 'Hubmap': {<br>\n        'kidney'        : 0.35,<br>\n        'prostate'      : 0.25,<br>\n        'largeintestine': 0.3,<br>\n        'spleen'        : 0.3,<br>\n        'lung'          : 0.1,<br>\n    },</p>\n<p><strong>Work</strong><br>\nSWA, TTA, Adjust the threshold, model fusion, Multi-input size fusion, Augmentation(Random flip, rotate, noise, HueSaturationValue, affine)<br>\n<strong>Not Work</strong><br>\n Augmentation - rgbtogbr and cutout,<br>\n External data and Stain normalization.</p>",
  "messages": [
    {
      "id": 1955404,
      "postDate": "2022-09-26T01:56:49.303Z",
      "content": "<p>First of all, thank the organizers for hosting this competition. <br>\nWe also thank <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> for his discussions and notebook.</p>\n<p><strong>Models</strong><br>\nCoat_small+daformer,size=768,folds=5,swa<br>\nCoat_small+daformer,size=768,folds=5,External data , swa <br>\nCoat_lite_medium+daformer,size=768,folds=5,swa<br>\nCoat_lite_medium+daformer,size=768,all data,swa<br>\nCoat_small(level5)+daformer_conv1*1,size=1152, folds=5, swa<br>\npvt_v2_b4+daformer_conv3x3,size=768,folds=5,swa<br>\npvt_v2_b4+daformer_conv3x3,size=768,all data,swa<br>\npvt_v2_b4+smp_unet,size=768,folds=5,swa<br>\npvt_v2_b4+smp_unet,size=768,all data,swa</p>\n<p><strong>Augmentations</strong><br>\nRandom flip, rotate, noise, HueSaturationValue, affine..</p>\n<p><strong>Data Processing</strong><br>\nWe have selected three methods of training data:<br>\nThe original data is divided into training data and validation data for training. <br>\nThe full data is trained. <br>\nThe external data is labeled and then added to the original data for training</p>\n<p><strong>Inference</strong><br>\nEnsemble<br>\nCoat_small+daformer:fold0 fold3<br>\nCoat_lite_medium+daformer: fold2 fold3 fold4 all_data<br>\nCoat_small(level5)+daformer_conv1*1 :fold0_1152<br>\npvt_v2_b4+daformer_conv3x3:fold1,fold2,fold3,fold4,all_data,<br>\nfold3_384,fold2_384,fold3_1024<br>\npvt_v2_b4+smp_unet:fold0,fold2,all_data</p>\n<p>TTA<br>\nFlip,Rorate</p>\n<p>Organ Threshold<br>\n 'Hubmap': {<br>\n        'kidney'        : 0.35,<br>\n        'prostate'      : 0.25,<br>\n        'largeintestine': 0.3,<br>\n        'spleen'        : 0.3,<br>\n        'lung'          : 0.1,<br>\n    },</p>\n<p><strong>Work</strong><br>\nSWA, TTA, Adjust the threshold, model fusion, Multi-input size fusion, Augmentation(Random flip, rotate, noise, HueSaturationValue, affine)<br>\n<strong>Not Work</strong><br>\n Augmentation - rgbtogbr and cutout,<br>\n External data and Stain normalization.</p>",
      "rawMarkdown": "First of all, thank the organizers for hosting this competition. \nWe also thank @hengck23 for his discussions and notebook.\n\n**Models**\nCoat_small+daformer,size=768,folds=5,swa\nCoat_small+daformer,size=768,folds=5,External data , swa \nCoat_lite_medium+daformer,size=768,folds=5,swa\nCoat_lite_medium+daformer,size=768,all data,swa\nCoat_small(level5)+daformer_conv1*1,size=1152, folds=5, swa\npvt_v2_b4+daformer_conv3x3,size=768,folds=5,swa\npvt_v2_b4+daformer_conv3x3,size=768,all data,swa\npvt_v2_b4+smp_unet,size=768,folds=5,swa\npvt_v2_b4+smp_unet,size=768,all data,swa\n\n**Augmentations**\nRandom flip, rotate, noise, HueSaturationValue, affine..\n\n**Data Processing**\nWe have selected three methods of training data:\nThe original data is divided into training data and validation data for training. \nThe full data is trained. \nThe external data is labeled and then added to the original data for training\n\n**Inference**\nEnsemble\nCoat_small+daformer:fold0 fold3\nCoat_lite_medium+daformer: fold2 fold3 fold4 all_data\nCoat_small(level5)+daformer_conv1*1 :fold0_1152\npvt_v2_b4+daformer_conv3x3:fold1,fold2,fold3,fold4,all_data,\nfold3_384,fold2_384,fold3_1024\npvt_v2_b4+smp_unet:fold0,fold2,all_data\n\nTTA\nFlip,Rorate\n\nOrgan Threshold\n 'Hubmap': {\n        'kidney'        : 0.35,\n        'prostate'      : 0.25,\n        'largeintestine': 0.3,\n        'spleen'        : 0.3,\n        'lung'          : 0.1,\n    },\n\n**Work**\nSWA, TTA, Adjust the threshold, model fusion, Multi-input size fusion, Augmentation(Random flip, rotate, noise, HueSaturationValue, affine)\n**Not Work**\n Augmentation - rgbtogbr and cutout,\n External data and Stain normalization.",
      "votes": 9
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1955404": "First of all, thank the organizers for hosting this competition. \nWe also thank @hengck23 for his discussions and notebook.\n\n**Models**\nCoat_small+daformer,size=768,folds=5,swa\nCoat_small+daformer,size=768,folds=5,External data , swa \nCoat_lite_medium+daformer,size=768,folds=5,swa\nCoat_lite_medium+daformer,size=768,all data,swa\nCoat_small(level5)+daformer_conv1*1,size=1152, folds=5, swa\npvt_v2_b4+daformer_conv3x3,size=768,folds=5,swa\npvt_v2_b4+daformer_conv3x3,size=768,all data,swa\npvt_v2_b4+smp_unet,size=768,folds=5,swa\npvt_v2_b4+smp_unet,size=768,all data,swa\n\n**Augmentations**\nRandom flip, rotate, noise, HueSaturationValue, affine..\n\n**Data Processing**\nWe have selected three methods of training data:\nThe original data is divided into training data and validation data for training. \nThe full data is trained. \nThe external data is labeled and then added to the original data for training\n\n**Inference**\nEnsemble\nCoat_small+daformer:fold0 fold3\nCoat_lite_medium+daformer: fold2 fold3 fold4 all_data\nCoat_small(level5)+daformer_conv1*1 :fold0_1152\npvt_v2_b4+daformer_conv3x3:fold1,fold2,fold3,fold4,all_data,\nfold3_384,fold2_384,fold3_1024\npvt_v2_b4+smp_unet:fold0,fold2,all_data\n\nTTA\nFlip,Rorate\n\nOrgan Threshold\n 'Hubmap': {\n        'kidney'        : 0.35,\n        'prostate'      : 0.25,\n        'largeintestine': 0.3,\n        'spleen'        : 0.3,\n        'lung'          : 0.1,\n    },\n\n**Work**\nSWA, TTA, Adjust the threshold, model fusion, Multi-input size fusion, Augmentation(Random flip, rotate, noise, HueSaturationValue, affine)\n**Not Work**\n Augmentation - rgbtogbr and cutout,\n External data and Stain normalization."
  }
}