{
  "id": 223209,
  "title": "Best single model",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/223209",
  "author_name": "Stanley Zheng",
  "post_date": "2021-03-02T23:37:26.938000",
  "votes": 26,
  "comment_count": 24,
  "views": 0,
  "content": "<p>Since this is a CSV competiton and likely a ton of the LB is ensembles, I'm curious to know your single model scores. I'll go first</p>\n<p>1 stage, 1 fold YOLOv4: 0.173<br>\nNMS labels CV 0.422 for mAP@0.5 w/TTA</p>\n<p>2 stage 5 fold YOLOv4: 0.248</p>\n<p>Using many tips from <a href=\"https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/213462\" target=\"_blank\">this great discussion</a></p>",
  "messages": [
    {
      "id": 1224641,
      "postDate": "2021-03-02T23:37:26.940Z",
      "content": "<p>Since this is a CSV competiton and likely a ton of the LB is ensembles, I'm curious to know your single model scores. I'll go first</p>\n<p>1 stage, 1 fold YOLOv4: 0.173<br>\nNMS labels CV 0.422 for mAP@0.5 w/TTA</p>\n<p>2 stage 5 fold YOLOv4: 0.248</p>\n<p>Using many tips from <a href=\"https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/213462\" target=\"_blank\">this great discussion</a></p>",
      "rawMarkdown": "Since this is a CSV competiton and likely a ton of the LB is ensembles, I'm curious to know your single model scores. I'll go first\n\n1 stage, 1 fold YOLOv4: 0.173\nNMS labels CV 0.422 for mAP@0.5 w/TTA\n\n2 stage 5 fold YOLOv4: 0.248\n\nUsing many tips from [this great discussion](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/213462)\n",
      "votes": 26
    },
    {
      "id": 1225640,
      "postDate": "2021-03-03T19:13:17.357Z",
      "content": "<p>All my results are for 1 fold.</p>\n<p>2 classes classifier -&gt; 0.993 AUC<br>\nDetection : VFNet on mmdetection -&gt; 0.357 for map@0.5</p>\n<p>LB: 0.259</p>",
      "rawMarkdown": "All my results are for 1 fold.\n\n2 classes classifier -> 0.993 AUC\nDetection : VFNet on mmdetection -> 0.357 for map@0.5\n\nLB: 0.259",
      "votes": 5,
      "replies": [
        {
          "id": 1225851,
          "postDate": "2021-03-04T02:16:44.320Z",
          "content": "<p>May I ask what is the score without <code>2 classes classifier</code>?</p>",
          "rawMarkdown": "May I ask what is the score without `2 classes classifier`?"
        },
        {
          "id": 1226003,
          "postDate": "2021-03-04T06:23:43.177Z",
          "content": "<p>VFNet predicts a lot of boxes, how did you filter or fuse them ? NMS, WBF ?</p>",
          "rawMarkdown": "VFNet predicts a lot of boxes, how did you filter or fuse them ? NMS, WBF ?"
        },
        {
          "id": 1226192,
          "postDate": "2021-03-04T10:15:58.220Z",
          "content": "<p>Without 2 classes it's 0.187.<br>\nWith MMdetection, NMS@0.6 is integrated in the test pipeline, I didn't change it </p>",
          "rawMarkdown": "Without 2 classes it's 0.187.\nWith MMdetection, NMS@0.6 is integrated in the test pipeline, I didn't change it ",
          "votes": 1
        },
        {
          "id": 1226693,
          "postDate": "2021-03-04T18:42:03.033Z",
          "content": "<p>Have you used nms or wbf on training stage ? Does your model predict a large number of bboxes ?</p>",
          "rawMarkdown": "Have you used nms or wbf on training stage ? Does your model predict a large number of bboxes ?"
        },
        {
          "id": 1226748,
          "postDate": "2021-03-04T20:01:53.437Z",
          "content": "<p><a href=\"https://www.kaggle.com/raufyagfarov\" target=\"_blank\">@raufyagfarov</a> No and not especially (~60/img) </p>",
          "rawMarkdown": "@raufyagfarov No and not especially (~60/img) ",
          "votes": 1
        },
        {
          "id": 1227038,
          "postDate": "2021-03-05T05:54:39.910Z",
          "content": "<p><a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> what is your single best classifier arch? Resnet? </p>",
          "rawMarkdown": "@matthieuplante what is your single best classifier arch? Resnet? "
        },
        {
          "id": 1227296,
          "postDate": "2021-03-05T11:51:46.543Z",
          "content": "<p><a href=\"https://www.kaggle.com/trushk\" target=\"_blank\">@trushk</a> the backbone of my arch is densenet </p>",
          "rawMarkdown": "@trushk the backbone of my arch is densenet ",
          "votes": 1
        },
        {
          "id": 1227492,
          "postDate": "2021-03-05T15:33:22.047Z",
          "content": "<p>Thanks <a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> </p>",
          "rawMarkdown": "Thanks @matthieuplante "
        },
        {
          "id": 1229005,
          "postDate": "2021-03-07T01:11:08.743Z",
          "content": "<p><a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> how many epochs for vfnet? Any extra augmentation?</p>",
          "rawMarkdown": "@matthieuplante how many epochs for vfnet? Any extra augmentation?"
        },
        {
          "id": 1229538,
          "postDate": "2021-03-07T12:21:48.187Z",
          "content": "<p>Hi! May i ask what is the image size you trained with?</p>",
          "rawMarkdown": "Hi! May i ask what is the image size you trained with?"
        },
        {
          "id": 1231334,
          "postDate": "2021-03-08T22:05:20.730Z",
          "content": "<p><a href=\"https://www.kaggle.com/angqx95\" target=\"_blank\">@angqx95</a> for the moment 1024x1024 but I'm working to use original images</p>",
          "rawMarkdown": "@angqx95 for the moment 1024x1024 but I'm working to use original images",
          "votes": 2
        },
        {
          "id": 1231373,
          "postDate": "2021-03-08T22:50:16.297Z",
          "content": "<p><a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> how do you plan on handling the variable sizes of the original images?</p>",
          "rawMarkdown": "@matthieuplante how do you plan on handling the variable sizes of the original images?"
        },
        {
          "id": 1231393,
          "postDate": "2021-03-08T23:52:39.630Z",
          "content": "<p>No idea yet haha</p>",
          "rawMarkdown": "No idea yet haha"
        },
        {
          "id": 1237434,
          "postDate": "2021-03-14T06:23:47.753Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1238328,
      "postDate": "2021-03-14T22:13:05.767Z",
      "content": "<p>Our best single model (1 fold) is 0.201. But we somehow couldn't improve it too much through 2 classes classifier yet.</p>",
      "rawMarkdown": "Our best single model (1 fold) is 0.201. But we somehow couldn't improve it too much through 2 classes classifier yet.",
      "votes": 1
    },
    {
      "id": 1225802,
      "postDate": "2021-03-04T00:30:32.147Z",
      "content": "<p>may I ask did you use any postprocessing? like add '14 1 0 0 1 1'</p>",
      "rawMarkdown": "may I ask did you use any postprocessing? like add '14 1 0 0 1 1'",
      "replies": [
        {
          "id": 1225803,
          "postDate": "2021-03-04T00:31:40.687Z",
          "content": "<p>No postprocessing</p>",
          "rawMarkdown": "No postprocessing"
        }
      ]
    },
    {
      "id": 1224779,
      "postDate": "2021-03-03T04:10:46.783Z",
      "content": "<p>Hi, thanks for sharing your result. Can i ask what ensemble technique do you use ?</p>",
      "rawMarkdown": "Hi, thanks for sharing your result. Can i ask what ensemble technique do you use ?",
      "replies": [
        {
          "id": 1224812,
          "postDate": "2021-03-03T04:55:32.927Z",
          "content": "<p>I'm just using NMS, haven't had time to tune it much</p>",
          "rawMarkdown": "I'm just using NMS, haven't had time to tune it much",
          "votes": 1
        },
        {
          "id": 1224960,
          "postDate": "2021-03-03T07:32:06.450Z",
          "content": "<p>Thank you </p>",
          "rawMarkdown": "Thank you "
        },
        {
          "id": 1225811,
          "postDate": "2021-03-04T00:57:20.427Z",
          "content": "<p>This is interesting. I just realized there are people doing research based on how to better ensemble boxes. I was just doing my own brute-force thingy.</p>",
          "rawMarkdown": "This is interesting. I just realized there are people doing research based on how to better ensemble boxes. I was just doing my own brute-force thingy.",
          "votes": 1
        },
        {
          "id": 1253234,
          "postDate": "2021-03-26T13:52:25.823Z",
          "content": "<p>Hello,why my scores decrease when ensembled?</p>",
          "rawMarkdown": "Hello,why my scores decrease when ensembled?"
        }
      ]
    },
    {
      "id": 1224900,
      "postDate": "2021-03-03T06:46:46.820Z",
      "rawMarkdown": "",
      "votes": -1,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1225640,
      "author_name": "Matthieu Planté",
      "author_url": "",
      "post_date": "2021-03-03T19:13:17.357000",
      "content": "<p>All my results are for 1 fold.</p>\n<p>2 classes classifier -&gt; 0.993 AUC<br>\nDetection : VFNet on mmdetection -&gt; 0.357 for map@0.5</p>\n<p>LB: 0.259</p>",
      "votes": 5,
      "replies": [
        {
          "id": 1225851,
          "author_name": "Phat Tran",
          "author_url": "",
          "post_date": "2021-03-04T02:16:44.320000",
          "content": "<p>May I ask what is the score without <code>2 classes classifier</code>?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1226003,
          "author_name": "Rauf  Yagfarov",
          "author_url": "",
          "post_date": "2021-03-04T06:23:43.177000",
          "content": "<p>VFNet predicts a lot of boxes, how did you filter or fuse them ? NMS, WBF ?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1226192,
          "author_name": "Matthieu Planté",
          "author_url": "",
          "post_date": "2021-03-04T10:15:58.220000",
          "content": "<p>Without 2 classes it's 0.187.<br>\nWith MMdetection, NMS@0.6 is integrated in the test pipeline, I didn't change it </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1226693,
          "author_name": "Rauf  Yagfarov",
          "author_url": "",
          "post_date": "2021-03-04T18:42:03.033000",
          "content": "<p>Have you used nms or wbf on training stage ? Does your model predict a large number of bboxes ?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1226748,
          "author_name": "Matthieu Planté",
          "author_url": "",
          "post_date": "2021-03-04T20:01:53.437000",
          "content": "<p><a href=\"https://www.kaggle.com/raufyagfarov\" target=\"_blank\">@raufyagfarov</a> No and not especially (~60/img) </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1227038,
          "author_name": "Trushant Kalyanpur",
          "author_url": "",
          "post_date": "2021-03-05T05:54:39.910000",
          "content": "<p><a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> what is your single best classifier arch? Resnet? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1227296,
          "author_name": "Matthieu Planté",
          "author_url": "",
          "post_date": "2021-03-05T11:51:46.543000",
          "content": "<p><a href=\"https://www.kaggle.com/trushk\" target=\"_blank\">@trushk</a> the backbone of my arch is densenet </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1227492,
          "author_name": "Trushant Kalyanpur",
          "author_url": "",
          "post_date": "2021-03-05T15:33:22.047000",
          "content": "<p>Thanks <a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1229005,
          "author_name": "Trushant Kalyanpur",
          "author_url": "",
          "post_date": "2021-03-07T01:11:08.743000",
          "content": "<p><a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> how many epochs for vfnet? Any extra augmentation?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1229538,
          "author_name": "aqx",
          "author_url": "",
          "post_date": "2021-03-07T12:21:48.187000",
          "content": "<p>Hi! May i ask what is the image size you trained with?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1231334,
          "author_name": "Matthieu Planté",
          "author_url": "",
          "post_date": "2021-03-08T22:05:20.730000",
          "content": "<p><a href=\"https://www.kaggle.com/angqx95\" target=\"_blank\">@angqx95</a> for the moment 1024x1024 but I'm working to use original images</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1231373,
          "author_name": "Daniel Shan",
          "author_url": "",
          "post_date": "2021-03-08T22:50:16.297000",
          "content": "<p><a href=\"https://www.kaggle.com/matthieuplante\" target=\"_blank\">@matthieuplante</a> how do you plan on handling the variable sizes of the original images?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1231393,
          "author_name": "Matthieu Planté",
          "author_url": "",
          "post_date": "2021-03-08T23:52:39.630000",
          "content": "<p>No idea yet haha</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1237434,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-03-14T06:23:47.753000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1238328,
      "author_name": "Fernando Camargo",
      "author_url": "",
      "post_date": "2021-03-14T22:13:05.767000",
      "content": "<p>Our best single model (1 fold) is 0.201. But we somehow couldn't improve it too much through 2 classes classifier yet.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1225802,
      "author_name": "Tian",
      "author_url": "",
      "post_date": "2021-03-04T00:30:32.147000",
      "content": "<p>may I ask did you use any postprocessing? like add '14 1 0 0 1 1'</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1225803,
          "author_name": "Stanley Zheng",
          "author_url": "",
          "post_date": "2021-03-04T00:31:40.687000",
          "content": "<p>No postprocessing</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1224779,
      "author_name": "Bùi Nhật Trường",
      "author_url": "",
      "post_date": "2021-03-03T04:10:46.783000",
      "content": "<p>Hi, thanks for sharing your result. Can i ask what ensemble technique do you use ?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1224812,
          "author_name": "Stanley Zheng",
          "author_url": "",
          "post_date": "2021-03-03T04:55:32.927000",
          "content": "<p>I'm just using NMS, haven't had time to tune it much</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1224960,
          "author_name": "Bùi Nhật Trường",
          "author_url": "",
          "post_date": "2021-03-03T07:32:06.450000",
          "content": "<p>Thank you </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1225811,
          "author_name": "Wilmer E. Henao",
          "author_url": "",
          "post_date": "2021-03-04T00:57:20.427000",
          "content": "<p>This is interesting. I just realized there are people doing research based on how to better ensemble boxes. I was just doing my own brute-force thingy.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1253234,
          "author_name": "HAaHAa",
          "author_url": "",
          "post_date": "2021-03-26T13:52:25.823000",
          "content": "<p>Hello,why my scores decrease when ensembled?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1224900,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-03-03T06:46:46.820000",
      "content": "",
      "votes": -1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1224641": "Since this is a CSV competiton and likely a ton of the LB is ensembles, I'm curious to know your single model scores. I'll go first\n\n1 stage, 1 fold YOLOv4: 0.173\nNMS labels CV 0.422 for mAP@0.5 w/TTA\n\n2 stage 5 fold YOLOv4: 0.248\n\nUsing many tips from [this great discussion](https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/213462)\n",
    "1225640": "All my results are for 1 fold.\n\n2 classes classifier -> 0.993 AUC\nDetection : VFNet on mmdetection -> 0.357 for map@0.5\n\nLB: 0.259",
    "1238328": "Our best single model (1 fold) is 0.201. But we somehow couldn't improve it too much through 2 classes classifier yet.",
    "1225802": "may I ask did you use any postprocessing? like add '14 1 0 0 1 1'",
    "1224779": "Hi, thanks for sharing your result. Can i ask what ensemble technique do you use ?",
    "1224900": ""
  }
}