{
  "id": 419114,
  "title": "Sharing model and LB",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/419114",
  "author_name": "dmitrykonovalov",
  "post_date": "2023-06-24T08:24:45.150000",
  "votes": 9,
  "comment_count": 15,
  "views": 0,
  "content": "<p>I am using detectron2 because I used it on many other projects. So my 1 fold trained on only ds1 gets LB0.401. I use TTA-flip and 4 sizes 512+64x1, + 64x2, +64x3, +64x4. I tried hflip and vflip, the same LB. <br>\nHas anyone got a good LB with yolov8 one model? What about just a standard unet+B0-B7? </p>",
  "messages": [
    {
      "id": 2315562,
      "postDate": "2023-06-24T08:24:45.150Z",
      "content": "<p>I am using detectron2 because I used it on many other projects. So my 1 fold trained on only ds1 gets LB0.401. I use TTA-flip and 4 sizes 512+64x1, + 64x2, +64x3, +64x4. I tried hflip and vflip, the same LB. <br>\nHas anyone got a good LB with yolov8 one model? What about just a standard unet+B0-B7? </p>",
      "rawMarkdown": "I am using detectron2 because I used it on many other projects. So my 1 fold trained on only ds1 gets LB0.401. I use TTA-flip and 4 sizes 512+64x1, + 64x2, +64x3, +64x4. I tried hflip and vflip, the same LB. \nHas anyone got a good LB with yolov8 one model? What about just a standard unet+B0-B7? ",
      "votes": 9
    },
    {
      "id": 2330237,
      "postDate": "2023-07-04T20:29:06.697Z",
      "content": "<p>Rtmdet large from mmdet 3.0.0rc5 + trainon both datset 1 and 2 (split 80:20) + image size 1024 + hard augmentations + flips as tta + dilation on test predctions - 0.460 LB</p>",
      "rawMarkdown": "Rtmdet large from mmdet 3.0.0rc5 + trainon both datset 1 and 2 (split 80:20) + image size 1024 + hard augmentations + flips as tta + dilation on test predctions - 0.460 LB",
      "votes": 2,
      "replies": [
        {
          "id": 2341867,
          "postDate": "2023-07-12T10:46:46.520Z",
          "content": "<p>Tiles are of size 512 x 512. Did you mesh tiles into a single big tiff and then split into 1024 x 1024? I am new to the competition :-)</p>",
          "rawMarkdown": "Tiles are of size 512 x 512. Did you mesh tiles into a single big tiff and then split into 1024 x 1024? I am new to the competition :-)"
        }
      ]
    },
    {
      "id": 2316496,
      "postDate": "2023-06-25T03:12:51.760Z",
      "content": "<p>It is impressive you reached 401 with single fold and only ds1. Can I ask which backbone and which training sizes you used ? <a href=\"https://www.kaggle.com/dmitrykonovalov\" target=\"_blank\">@dmitrykonovalov</a> </p>",
      "rawMarkdown": "It is impressive you reached 401 with single fold and only ds1. Can I ask which backbone and which training sizes you used ? @dmitrykonovalov ",
      "replies": [
        {
          "id": 2317108,
          "postDate": "2023-06-25T12:52:04.653Z",
          "content": "<p>cfg.model_det2_name = 'COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml'<br>\nbase_min_size = 512<br>\ncfg.INPUT_MIN_SIZE_TRAIN = [base_min_size - base_min_size // 3, base_min_size + base_min_size // 2]<br>\ncfg.INPUT_MIN_SIZE_TRAIN_SAMPLING = \"range\"  # \"choice\" # \"range\"<br>\ncfg.INPUT_MAX_SIZE_TRAIN = int(base_min_size * 1.5)</p>",
          "rawMarkdown": "cfg.model_det2_name = 'COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml'\nbase_min_size = 512\ncfg.INPUT_MIN_SIZE_TRAIN = [base_min_size - base_min_size // 3, base_min_size + base_min_size // 2]\ncfg.INPUT_MIN_SIZE_TRAIN_SAMPLING = \"range\"  # \"choice\" # \"range\"\ncfg.INPUT_MAX_SIZE_TRAIN = int(base_min_size * 1.5)",
          "votes": 2
        }
      ]
    },
    {
      "id": 2315827,
      "postDate": "2023-06-24T13:00:16.073Z",
      "content": "<p>Hello, could you explain more how you use the 4 different sizes?</p>",
      "rawMarkdown": "Hello, could you explain more how you use the 4 different sizes?",
      "replies": [
        {
          "id": 2317102,
          "postDate": "2023-06-25T12:46:53.653Z",
          "content": "<p>predictor = DefaultPredictor(det2_cfg)<br>\ndet2_cfg.TEST.AUG.ENABLED = True<br>\ndet2_cfg.TEST.AUG.FLIP = True<br>\ndet2_cfg.TEST.AUG.MIN_SIZES = (512, 512+64, ). # etc<br>\ndet2_cfg.TEST.AUG.MAX_SIZE = 1024<br>\nmodel = GeneralizedRCNNWithTTA(det2_cfg, predictor.model, tta_mapper=None, batch_size=1)</p>",
          "rawMarkdown": "predictor = DefaultPredictor(det2_cfg)\ndet2_cfg.TEST.AUG.ENABLED = True\ndet2_cfg.TEST.AUG.FLIP = True\ndet2_cfg.TEST.AUG.MIN_SIZES = (512, 512+64, ). # etc\ndet2_cfg.TEST.AUG.MAX_SIZE = 1024\nmodel = GeneralizedRCNNWithTTA(det2_cfg, predictor.model, tta_mapper=None, batch_size=1)",
          "votes": 2
        }
      ]
    },
    {
      "id": 2315716,
      "postDate": "2023-06-24T10:48:39.227Z",
      "content": "<p>What are you using for cv if you are training on full ds1?</p>",
      "rawMarkdown": "What are you using for cv if you are training on full ds1?",
      "replies": [
        {
          "id": 2317082,
          "postDate": "2023-06-25T12:34:43.783Z",
          "content": "<p>just 5-fold split</p>",
          "rawMarkdown": "just 5-fold split",
          "votes": 1
        }
      ]
    },
    {
      "id": 2315695,
      "postDate": "2023-06-24T10:25:48.563Z",
      "content": "<p>did you use dilation in LB0.401?</p>",
      "rawMarkdown": "did you use dilation in LB0.401?",
      "replies": [
        {
          "id": 2317088,
          "postDate": "2023-06-25T12:39:57.170Z",
          "content": "<p>trying now</p>",
          "rawMarkdown": "trying now"
        },
        {
          "id": 2318356,
          "postDate": "2023-06-26T10:16:23.260Z",
          "content": "<p>no difference in my case</p>",
          "rawMarkdown": "no difference in my case",
          "replies": [
            {
              "id": 2318623,
              "postDate": "2023-06-26T13:33:23.727Z",
              "content": "<p>it's too strange, many reported score improved after dilation, including me, the increase is about 0.1</p>",
              "rawMarkdown": "it's too strange, many reported score improved after dilation, including me, the increase is about 0.1"
            },
            {
              "id": 2333480,
              "postDate": "2023-07-07T02:56:37.840Z",
              "content": "<p>I think this is because he is only using ds1 as training dataset</p>",
              "rawMarkdown": "I think this is because he is only using ds1 as training dataset",
              "votes": 1
            },
            {
              "id": 2336537,
              "postDate": "2023-07-09T12:25:31.467Z",
              "content": "<p>Is it possible to train a good model with only 416 images in ds1?</p>",
              "rawMarkdown": "\nIs it possible to train a good model with only 416 images in ds1?"
            }
          ]
        }
      ]
    },
    {
      "id": 2321152,
      "postDate": "2023-06-28T10:18:21.140Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 2330237,
      "author_name": "Kostiantyn Maksymov",
      "author_url": "",
      "post_date": "2023-07-04T20:29:06.697000",
      "content": "<p>Rtmdet large from mmdet 3.0.0rc5 + trainon both datset 1 and 2 (split 80:20) + image size 1024 + hard augmentations + flips as tta + dilation on test predctions - 0.460 LB</p>",
      "votes": 2,
      "replies": [
        {
          "id": 2341867,
          "author_name": "GUNER",
          "author_url": "",
          "post_date": "2023-07-12T10:46:46.520000",
          "content": "<p>Tiles are of size 512 x 512. Did you mesh tiles into a single big tiff and then split into 1024 x 1024? I am new to the competition :-)</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2316496,
      "author_name": "Liam Nguyen",
      "author_url": "",
      "post_date": "2023-06-25T03:12:51.760000",
      "content": "<p>It is impressive you reached 401 with single fold and only ds1. Can I ask which backbone and which training sizes you used ? <a href=\"https://www.kaggle.com/dmitrykonovalov\" target=\"_blank\">@dmitrykonovalov</a> </p>",
      "votes": 0,
      "replies": [
        {
          "id": 2317108,
          "author_name": "dmitrykonovalov",
          "author_url": "",
          "post_date": "2023-06-25T12:52:04.653000",
          "content": "<p>cfg.model_det2_name = 'COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml'<br>\nbase_min_size = 512<br>\ncfg.INPUT_MIN_SIZE_TRAIN = [base_min_size - base_min_size // 3, base_min_size + base_min_size // 2]<br>\ncfg.INPUT_MIN_SIZE_TRAIN_SAMPLING = \"range\"  # \"choice\" # \"range\"<br>\ncfg.INPUT_MAX_SIZE_TRAIN = int(base_min_size * 1.5)</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 2315827,
      "author_name": "Heitor Rapela Medeiros",
      "author_url": "",
      "post_date": "2023-06-24T13:00:16.073000",
      "content": "<p>Hello, could you explain more how you use the 4 different sizes?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 2317102,
          "author_name": "dmitrykonovalov",
          "author_url": "",
          "post_date": "2023-06-25T12:46:53.653000",
          "content": "<p>predictor = DefaultPredictor(det2_cfg)<br>\ndet2_cfg.TEST.AUG.ENABLED = True<br>\ndet2_cfg.TEST.AUG.FLIP = True<br>\ndet2_cfg.TEST.AUG.MIN_SIZES = (512, 512+64, ). # etc<br>\ndet2_cfg.TEST.AUG.MAX_SIZE = 1024<br>\nmodel = GeneralizedRCNNWithTTA(det2_cfg, predictor.model, tta_mapper=None, batch_size=1)</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 2315716,
      "author_name": "Camilla",
      "author_url": "",
      "post_date": "2023-06-24T10:48:39.227000",
      "content": "<p>What are you using for cv if you are training on full ds1?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 2317082,
          "author_name": "dmitrykonovalov",
          "author_url": "",
          "post_date": "2023-06-25T12:34:43.783000",
          "content": "<p>just 5-fold split</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 2315695,
      "author_name": "Yi Wu",
      "author_url": "",
      "post_date": "2023-06-24T10:25:48.563000",
      "content": "<p>did you use dilation in LB0.401?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 2317088,
          "author_name": "dmitrykonovalov",
          "author_url": "",
          "post_date": "2023-06-25T12:39:57.170000",
          "content": "<p>trying now</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2318356,
          "author_name": "dmitrykonovalov",
          "author_url": "",
          "post_date": "2023-06-26T10:16:23.260000",
          "content": "<p>no difference in my case</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2318623,
              "author_name": "Yi Wu",
              "author_url": "",
              "post_date": "2023-06-26T13:33:23.727000",
              "content": "<p>it's too strange, many reported score improved after dilation, including me, the increase is about 0.1</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2333480,
              "author_name": "Feng Qilong",
              "author_url": "",
              "post_date": "2023-07-07T02:56:37.840000",
              "content": "<p>I think this is because he is only using ds1 as training dataset</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 2336537,
              "author_name": "fanshutou",
              "author_url": "",
              "post_date": "2023-07-09T12:25:31.467000",
              "content": "<p>Is it possible to train a good model with only 416 images in ds1?</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2321152,
      "author_name": "",
      "author_url": "",
      "post_date": "2023-06-28T10:18:21.140000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2315562": "I am using detectron2 because I used it on many other projects. So my 1 fold trained on only ds1 gets LB0.401. I use TTA-flip and 4 sizes 512+64x1, + 64x2, +64x3, +64x4. I tried hflip and vflip, the same LB. \nHas anyone got a good LB with yolov8 one model? What about just a standard unet+B0-B7? ",
    "2330237": "Rtmdet large from mmdet 3.0.0rc5 + trainon both datset 1 and 2 (split 80:20) + image size 1024 + hard augmentations + flips as tta + dilation on test predctions - 0.460 LB",
    "2316496": "It is impressive you reached 401 with single fold and only ds1. Can I ask which backbone and which training sizes you used ? @dmitrykonovalov ",
    "2315827": "Hello, could you explain more how you use the 4 different sizes?",
    "2315716": "What are you using for cv if you are training on full ds1?",
    "2315695": "did you use dilation in LB0.401?",
    "2321152": ""
  }
}