{
  "id": 429873,
  "title": "5th place solution. Higher Resolution and Dataset1 is all you need.",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/writeups/zw-5th-place-solution-higher-resolution-and-datase",
  "author_name": "",
  "post_date": "2023-08-07T14:39:31.610088500Z",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Thank you to the organizers for hosting this competition！<br>\nThank you to all the people who helped me!  I have learned a lot from you！<br>\nThis is my first time winning a gold medal！I am a master finally！</p>\n<h1>data</h1>\n<p>I only use dataset1 in my final submission，with coco pretrain.<br>\n5folds splits were used in local cv</p>\n<h1>Training</h1>\n<h2>Model</h2>\n<p>I tried detectoRS-r101, htc-db-b. Htc-db-b performs better in both local cv（score 0.439） and lb（score 0.570) .</p>\n<h2>Training Strategy</h2>\n<p>I trained with mmdet2, default 1x schedule and 1x swa training. <br>\nImage scales were set as (512，512)-(1536,1536).<br>\nRandomFlip and Rotate90 aug were used.</p>\n<h1>Inference</h1>\n<p>TTA i used in inference. TTA scales is [(640,640),(800,800),(1024,1024),(1408,1408),(1696,1696),(1920,1920)] while TTA flip_direction is [horizontal,vertical].<br>\nReplace rcnn'nms by weighted cluster-nms with DIoU.</p>\n<h1>Final Submission</h1>\n<p>My final submission was 1 fold htc-db-b，trained in dataset1.</p>\n<h1>Others</h1>\n<h2>Ensemble</h2>\n<p>I tried ensemble of different folds，which decrease both cv and public score，so I didn't chose it as my final submission，but it increase private score finally. Only 2folds ensemble increase private from 0.57 to 0.573.</p>\n<h2>Larger Model</h2>\n<p>I tired htc-db-l，which local cv and public socre both lower than htc-db-b，but have higher score（0.573）in private score.<br>\n ## Tried but not worked<br>\ndata augmentation: copy-paste，mixup，HSV，cutout…<br>\ndataset23: I tried pretrained in dataset2 then fintune in dataset1, which decrease score both in cv and lb. I tried pseudo label, which also didn’t work for me.</p>",
  "messages": [
    {
      "id": "2378229",
      "postDate": "08/07/2023 14:39:31",
      "content": "<p>Thank you to the organizers for hosting this competition！<br>\nThank you to all the people who helped me!  I have learned a lot from you！<br>\nThis is my first time winning a gold medal！I am a master finally！</p>\n<h1>data</h1>\n<p>I only use dataset1 in my final submission，with coco pretrain.<br>\n5folds splits were used in local cv</p>\n<h1>Training</h1>\n<h2>Model</h2>\n<p>I tried detectoRS-r101, htc-db-b. Htc-db-b performs better in both local cv（score 0.439） and lb（score 0.570) .</p>\n<h2>Training Strategy</h2>\n<p>I trained with mmdet2, default 1x schedule and 1x swa training. <br>\nImage scales were set as (512，512)-(1536,1536).<br>\nRandomFlip and Rotate90 aug were used.</p>\n<h1>Inference</h1>\n<p>TTA i used in inference. TTA scales is [(640,640),(800,800),(1024,1024),(1408,1408),(1696,1696),(1920,1920)] while TTA flip_direction is [horizontal,vertical].<br>\nReplace rcnn'nms by weighted cluster-nms with DIoU.</p>\n<h1>Final Submission</h1>\n<p>My final submission was 1 fold htc-db-b，trained in dataset1.</p>\n<h1>Others</h1>\n<h2>Ensemble</h2>\n<p>I tried ensemble of different folds，which decrease both cv and public score，so I didn't chose it as my final submission，but it increase private score finally. Only 2folds ensemble increase private from 0.57 to 0.573.</p>\n<h2>Larger Model</h2>\n<p>I tired htc-db-l，which local cv and public socre both lower than htc-db-b，but have higher score（0.573）in private score.<br>\n ## Tried but not worked<br>\ndata augmentation: copy-paste，mixup，HSV，cutout…<br>\ndataset23: I tried pretrained in dataset2 then fintune in dataset1, which decrease score both in cv and lb. I tried pseudo label, which also didn’t work for me.</p>",
      "rawMarkdown": "Thank you to the organizers for hosting this competition！\nThank you to all the people who helped me!  I have learned a lot from you！\nThis is my first time winning a gold medal！I am a master finally！\n\n\n# data\nI only use dataset1 in my final submission，with coco pretrain.\n5folds splits were used in local cv\n\n# Training\n## Model\nI tried detectoRS-r101, htc-db-b. Htc-db-b performs better in both local cv（score 0.439） and lb（score 0.570) .\n\n## Training Strategy\nI trained with mmdet2, default 1x schedule and 1x swa training. \nImage scales were set as (512，512)-(1536,1536).\nRandomFlip and Rotate90 aug were used.\n\n# Inference\nTTA i used in inference. TTA scales is [(640,640),(800,800),(1024,1024),(1408,1408),(1696,1696),(1920,1920)] while TTA flip_direction is [horizontal,vertical].\nReplace rcnn'nms by weighted cluster-nms with DIoU.\n\n# Final Submission\nMy final submission was 1 fold htc-db-b，trained in dataset1.\n\n# Others\n## Ensemble\nI tried ensemble of different folds，which decrease both cv and public score，so I didn't chose it as my final submission，but it increase private score finally. Only 2folds ensemble increase private from 0.57 to 0.573.\n## Larger Model\nI tired htc-db-l，which local cv and public socre both lower than htc-db-b，but have higher score（0.573）in private score.\n ## Tried but not worked\ndata augmentation: copy-paste，mixup，HSV，cutout...\ndataset23: I tried pretrained in dataset2 then fintune in dataset1, which decrease score both in cv and lb. I tried pseudo label, which also didn’t work for me.",
      "votes": null
    },
    {
      "id": "2382059",
      "postDate": "08/09/2023 14:56:52",
      "content": "<p>inference code:   <a href=\"https://www.kaggle.com/code/wushaodong/5th-solution-htc-db-b-single-fold-submission\" target=\"_blank\">https://www.kaggle.com/code/wushaodong/5th-solution-htc-db-b-single-fold-submission</a></p>",
      "rawMarkdown": "inference code:   https://www.kaggle.com/code/wushaodong/5th-solution-htc-db-b-single-fold-submission",
      "votes": null
    },
    {
      "id": "2420114",
      "postDate": "09/02/2023 12:25:12",
      "content": "<p>have you tried dilate or change the mask binary threshold?</p>",
      "rawMarkdown": "have you tried dilate or change the mask binary threshold?",
      "votes": null
    },
    {
      "id": "2422373",
      "postDate": "09/04/2023 01:01:06",
      "content": "<p>both not   </p>",
      "rawMarkdown": "both not",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2382059,
      "author_name": "wushaodong",
      "author_url": "",
      "post_date": "08/09/2023 14:56:52",
      "content": "<p>inference code:   <a href=\"https://www.kaggle.com/code/wushaodong/5th-solution-htc-db-b-single-fold-submission\" target=\"_blank\">https://www.kaggle.com/code/wushaodong/5th-solution-htc-db-b-single-fold-submission</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2420114,
      "author_name": "clwwlc",
      "author_url": "",
      "post_date": "09/02/2023 12:25:12",
      "content": "<p>have you tried dilate or change the mask binary threshold?</p>",
      "votes": null,
      "replies": [
        {
          "id": 2422373,
          "author_name": "wushaodong",
          "author_url": "",
          "post_date": "09/04/2023 01:01:06",
          "content": "<p>both not   </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2378229": "Thank you to the organizers for hosting this competition！\nThank you to all the people who helped me!  I have learned a lot from you！\nThis is my first time winning a gold medal！I am a master finally！\n\n\n# data\nI only use dataset1 in my final submission，with coco pretrain.\n5folds splits were used in local cv\n\n# Training\n## Model\nI tried detectoRS-r101, htc-db-b. Htc-db-b performs better in both local cv（score 0.439） and lb（score 0.570) .\n\n## Training Strategy\nI trained with mmdet2, default 1x schedule and 1x swa training. \nImage scales were set as (512，512)-(1536,1536).\nRandomFlip and Rotate90 aug were used.\n\n# Inference\nTTA i used in inference. TTA scales is [(640,640),(800,800),(1024,1024),(1408,1408),(1696,1696),(1920,1920)] while TTA flip_direction is [horizontal,vertical].\nReplace rcnn'nms by weighted cluster-nms with DIoU.\n\n# Final Submission\nMy final submission was 1 fold htc-db-b，trained in dataset1.\n\n# Others\n## Ensemble\nI tried ensemble of different folds，which decrease both cv and public score，so I didn't chose it as my final submission，but it increase private score finally. Only 2folds ensemble increase private from 0.57 to 0.573.\n## Larger Model\nI tired htc-db-l，which local cv and public socre both lower than htc-db-b，but have higher score（0.573）in private score.\n ## Tried but not worked\ndata augmentation: copy-paste，mixup，HSV，cutout...\ndataset23: I tried pretrained in dataset2 then fintune in dataset1, which decrease score both in cv and lb. I tried pseudo label, which also didn’t work for me.",
    "2382059": "inference code:   https://www.kaggle.com/code/wushaodong/5th-solution-htc-db-b-single-fold-submission",
    "2420114": "have you tried dilate or change the mask binary threshold?",
    "2422373": "both not"
  },
  "source": "meta"
}