{
  "id": 100985,
  "title": "Classification scores",
  "url": "/competitions/aptos2019-blindness-detection/discussion/100985",
  "author_name": "Miroslav Valan",
  "post_date": "2019-07-22T13:22:58.230000",
  "votes": 17,
  "comment_count": 34,
  "views": 0,
  "content": "<p>There is another topic where teams share their scores and CV vs LB correlations. It seems that most of the top teams use regression. Thus, I am creating this separate topic for those who did some experiments with classification and are willing to share their experience using (1) official data, (2) external data and (3) correlation with regression.</p>\n\n<ol>\n<li>0.789 official train set (5fold). - 0.778 single fold</li>\n<li>0.799 with external (5fold after single trial)</li>\n<li>haven't tried yet - will give it a shot soon</li>\n</ol>\n\n<p>UPDATE 2019-07-23:\n1. 0.795 official data only; classification</p>\n\n<p>UPDATE 2019-08-23:\n1. 0.817 official data only; simple classification</p>",
  "messages": [
    {
      "id": 581851,
      "postDate": "2019-07-22T13:22:58.230Z",
      "content": "<p>There is another topic where teams share their scores and CV vs LB correlations. It seems that most of the top teams use regression. Thus, I am creating this separate topic for those who did some experiments with classification and are willing to share their experience using (1) official data, (2) external data and (3) correlation with regression.</p>\n\n<ol>\n<li>0.789 official train set (5fold). - 0.778 single fold</li>\n<li>0.799 with external (5fold after single trial)</li>\n<li>haven't tried yet - will give it a shot soon</li>\n</ol>\n\n<p>UPDATE 2019-07-23:\n1. 0.795 official data only; classification</p>\n\n<p>UPDATE 2019-08-23:\n1. 0.817 official data only; simple classification</p>",
      "rawMarkdown": "There is another topic where teams share their scores and CV vs LB correlations. It seems that most of the top teams use regression. Thus, I am creating this separate topic for those who did some experiments with classification and are willing to share their experience using (1) official data, (2) external data and (3) correlation with regression.\n\n1. 0.789 official train set (5fold). - 0.778 single fold\n2. 0.799 with external (5fold after single trial)\n3. haven't tried yet - will give it a shot soon\n\nUPDATE 2019-07-23:\n1. 0.795 official data only; classification\n\nUPDATE 2019-08-23:\n1. 0.817 official data only; simple classification\n",
      "votes": 16
    },
    {
      "id": 616386,
      "postDate": "2019-09-03T05:07:07.907Z",
      "content": "<p>I'm also doing only classification, so .834 is possible with an ensemble :)\n(pre-trained on 2015, fine tune on 2019, in total 4 model weights - 2x B4, 1x B3, 1x resnext, TTA, no pseudo labelling yet)</p>",
      "rawMarkdown": "I'm also doing only classification, so .834 is possible with an ensemble :)\n(pre-trained on 2015, fine tune on 2019, in total 4 model weights - 2x B4, 1x B3, 1x resnext, TTA, no pseudo labelling yet)",
      "votes": 1,
      "replies": [
        {
          "id": 639614,
          "postDate": "2019-10-03T12:04:32.790Z",
          "content": "<p>is your final submission based on classification only?</p>",
          "rawMarkdown": "is your final submission based on classification only?"
        },
        {
          "id": 643335,
          "postDate": "2019-10-07T12:22:20.587Z",
          "content": "<p>yes, I was using only class. models - ordinary classification - i.e. [11100]\ndon't have the code in my head but I think I used a \"multiplier\" of 1.8 - i.e. difference of 1 class would be penalized with 1.8, 2 classes 1.8^2 etc.</p>",
          "rawMarkdown": "yes, I was using only class. models - ordinary classification - i.e. [11100]\ndon't have the code in my head but I think I used a \"multiplier\" of 1.8 - i.e. difference of 1 class would be penalized with 1.8, 2 classes 1.8^2 etc."
        }
      ]
    },
    {
      "id": 582484,
      "postDate": "2019-07-23T08:26:38.133Z",
      "content": "<p>0.739 - ResNet 50 - official train data - classification - single fold.</p>",
      "rawMarkdown": "0.739 - ResNet 50 - official train data - classification - single fold.",
      "votes": 1
    },
    {
      "id": 582415,
      "postDate": "2019-07-23T06:46:15.733Z",
      "content": "<p>Not doing especially great - .689 using ResNext, external data, kappa optimization, but no TTA or CV (classification only). Some pointers would be greatly appreciated.</p>",
      "rawMarkdown": "Not doing especially great - .689 using ResNext, external data, kappa optimization, but no TTA or CV (classification only). Some pointers would be greatly appreciated.",
      "votes": 1,
      "replies": [
        {
          "id": 582416,
          "postDate": "2019-07-23T06:46:57.713Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 582069,
      "postDate": "2019-07-22T18:08:25.833Z",
      "content": "<p>My current score (0.786, 5fold, with external) is classification only</p>",
      "rawMarkdown": "My current score (0.786, 5fold, with external) is classification only",
      "votes": 1,
      "replies": [
        {
          "id": 582705,
          "postDate": "2019-07-23T13:41:19.787Z",
          "content": "<p>Thx, how much boost you got with external. I did only one trial and it added 0.025 to my setup at the time. (.774---&gt;.799</p>",
          "rawMarkdown": "Thx, how much boost you got with external. I did only one trial and it added 0.025 to my setup at the time. (.774---&gt;.799"
        },
        {
          "id": 582989,
          "postDate": "2019-07-23T21:21:15.510Z",
          "content": "<p>I think I got from 0.705 to 0.760, single fold</p>",
          "rawMarkdown": "I think I got from 0.705 to 0.760, single fold",
          "votes": 1
        },
        {
          "id": 583222,
          "postDate": "2019-07-24T07:16:33.883Z",
          "content": "<p>That's encouraging.</p>",
          "rawMarkdown": "That's encouraging."
        }
      ]
    },
    {
      "id": 585266,
      "postDate": "2019-07-27T07:28:53.820Z",
      "content": "<p>Hi，did you use the ben's preprocessing?</p>",
      "rawMarkdown": "Hi，did you use the ben's preprocessing?",
      "votes": -1,
      "replies": [
        {
          "id": 585368,
          "postDate": "2019-07-27T10:22:28.737Z",
          "content": "<p>I did not</p>",
          "rawMarkdown": "I did not"
        },
        {
          "id": 585449,
          "postDate": "2019-07-27T13:34:25.907Z",
          "content": "<p>yep, thank you, but I think the deep model is easy to overfitting for this competition.</p>",
          "rawMarkdown": "yep, thank you, but I think the deep model is easy to overfitting for this competition."
        }
      ]
    },
    {
      "id": 617106,
      "postDate": "2019-09-03T18:41:38.810Z",
      "content": "<p><a href=\"/valanm\">@valanm</a> excellent single-model classification results with competition data alone. Do you mind sharing which model architecture you are using? Efficient nets or other? Also, your ensemble seems to get you 0.83+, which 3 models are you using for that? Are these all trained on competition data only?\nI am getting 0.77 with efficient b3 with image size 300 using this competition data only. Trying to hit at least 0.8 with competition data, but struggling...</p>",
      "rawMarkdown": "@valanm excellent single-model classification results with competition data alone. Do you mind sharing which model architecture you are using? Efficient nets or other? Also, your ensemble seems to get you 0.83+, which 3 models are you using for that? Are these all trained on competition data only?\nI am getting 0.77 with efficient b3 with image size 300 using this competition data only. Trying to hit at least 0.8 with competition data, but struggling...",
      "replies": [
        {
          "id": 622871,
          "postDate": "2019-09-10T07:23:29.733Z",
          "content": "<p>Please read my solution</p>",
          "rawMarkdown": "Please read my solution"
        }
      ]
    },
    {
      "id": 616377,
      "postDate": "2019-09-03T04:37:17.533Z",
      "content": "<p>Update: 0.831 average of 3 classification models. </p>",
      "rawMarkdown": "Update: 0.831 average of 3 classification models. \n\n",
      "replies": [
        {
          "id": 617094,
          "postDate": "2019-09-03T18:35:19.707Z",
          "content": "<p>You seem to be getting pretty high scores using only the competition data with classification models. Is this ensemble based on competition data only as well, or does this include models pre-trained on old competition data?\nAre using TTA?</p>",
          "rawMarkdown": "You seem to be getting pretty high scores using only the competition data with classification models. Is this ensemble based on competition data only as well, or does this include models pre-trained on old competition data?\nAre using TTA?"
        }
      ]
    },
    {
      "id": 612788,
      "postDate": "2019-08-30T02:20:08.323Z",
      "content": "<p>Hi, do you guys change to multi class label? \nexp:  label 4 == [1 1 1 1 0]\n         label 5 == [1 1 1 1 1]\nI get a bad score 0.726, by using ben's crop, multi lables, classification, kappa score.  </p>",
      "rawMarkdown": "Hi, do you guys change to multi class label? \nexp:  label 4 == [1 1 1 1 0]\n         label 5 == [1 1 1 1 1]\nI get a bad score 0.726, by using ben's crop, multi lables, classification, kappa score.  ",
      "replies": [
        {
          "id": 612911,
          "postDate": "2019-08-30T04:49:48.917Z",
          "content": "<p>No</p>",
          "rawMarkdown": "No"
        },
        {
          "id": 612974,
          "postDate": "2019-08-30T06:06:58.800Z",
          "content": "<p>Thanks 😬 </p>",
          "rawMarkdown": "Thanks 😬 "
        }
      ]
    },
    {
      "id": 606053,
      "postDate": "2019-08-23T06:13:31.517Z",
      "content": "<p>0.817 official data only; simple classification</p>",
      "rawMarkdown": " 0.817 official data only; simple classification",
      "replies": [
        {
          "id": 606069,
          "postDate": "2019-08-23T06:40:51.577Z",
          "content": "<p>May I ask what is your highest accuracy rate base on CV?</p>",
          "rawMarkdown": "May I ask what is your highest accuracy rate base on CV?"
        },
        {
          "id": 606214,
          "postDate": "2019-08-23T10:27:14.247Z",
          "content": "<p>0.925+ kappa 0.86+ accuracy</p>",
          "rawMarkdown": "0.925+ kappa 0.86+ accuracy",
          "votes": 1
        },
        {
          "id": 606223,
          "postDate": "2019-08-23T10:45:39.247Z",
          "content": "<p>Hi, what model did you use? and 5 fold?</p>",
          "rawMarkdown": "Hi, what model did you use? and 5 fold?"
        }
      ]
    },
    {
      "id": 584369,
      "postDate": "2019-07-25T20:38:29.533Z",
      "content": "<p>This may sound a little silly but are you guys treating this as a plain old classification task with one hot vector outputs or as a multilabel classification task like <a href=\"https://www.kaggle.com/lextoumbourou/blindness-detection-resnet34-ordinal-targets\">this</a> one?</p>",
      "rawMarkdown": "This may sound a little silly but are you guys treating this as a plain old classification task with one hot vector outputs or as a multilabel classification task like [this](https://www.kaggle.com/lextoumbourou/blindness-detection-resnet34-ordinal-targets) one?",
      "replies": [
        {
          "id": 584373,
          "postDate": "2019-07-25T20:44:14.627Z",
          "content": "<p>plain old classification task in my case :) until  this moment </p>",
          "rawMarkdown": "plain old classification task in my case :) until  this moment ",
          "votes": 1
        }
      ]
    },
    {
      "id": 583677,
      "postDate": "2019-07-24T19:56:03.133Z",
      "content": "<p>What architecture are you using?</p>",
      "rawMarkdown": "What architecture are you using?",
      "replies": [
        {
          "id": 583700,
          "postDate": "2019-07-24T21:03:42.827Z",
          "content": "<p>resnext101</p>",
          "rawMarkdown": "resnext101",
          "votes": 1
        },
        {
          "id": 583707,
          "postDate": "2019-07-24T21:20:11.893Z",
          "content": "<p>Ok are you using a lot of different tricks, or more or less out-of-the-box? </p>",
          "rawMarkdown": "Ok are you using a lot of different tricks, or more or less out-of-the-box? "
        },
        {
          "id": 584379,
          "postDate": "2019-07-25T20:59:47.210Z",
          "content": "<p>i do use some minor tricks compared to default settings but nothing that can't be seen in public kernels</p>",
          "rawMarkdown": " i do use some minor tricks compared to default settings but nothing that can't be seen in public kernels"
        },
        {
          "id": 584410,
          "postDate": "2019-07-25T22:26:23.397Z",
          "content": "<p>Nice! Thanks!</p>",
          "rawMarkdown": "Nice! Thanks!"
        }
      ]
    },
    {
      "id": 583601,
      "postDate": "2019-07-24T17:26:06.910Z",
      "content": "<p><code>0.795 official data only; classification</code> </p>\n\n<p>Thank you for sharing the information, you can learn more.\nDid you get this result using the classification into 5 classes? Without using regression?</p>",
      "rawMarkdown": "`0.795 official data only; classification` \n\nThank you for sharing the information, you can learn more.\nDid you get this result using the classification into 5 classes? Without using regression?",
      "replies": [
        {
          "id": 583602,
          "postDate": "2019-07-24T17:29:54.530Z",
          "content": "<p>Yes. Only classification trained on official data</p>",
          "rawMarkdown": "Yes. Only classification trained on official data",
          "votes": 1
        },
        {
          "id": 583608,
          "postDate": "2019-07-24T17:41:41.450Z",
          "content": "<p>Thanks for the answer. I hope you make your result even better, good luck</p>",
          "rawMarkdown": "Thanks for the answer. I hope you make your result even better, good luck"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 616386,
      "author_name": "steelrose",
      "author_url": "",
      "post_date": "2019-09-03T05:07:07.907000",
      "content": "<p>I'm also doing only classification, so .834 is possible with an ensemble :)\n(pre-trained on 2015, fine tune on 2019, in total 4 model weights - 2x B4, 1x B3, 1x resnext, TTA, no pseudo labelling yet)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 639614,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-10-03T12:04:32.790000",
          "content": "<p>is your final submission based on classification only?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 643335,
          "author_name": "steelrose",
          "author_url": "",
          "post_date": "2019-10-07T12:22:20.587000",
          "content": "<p>yes, I was using only class. models - ordinary classification - i.e. [11100]\ndon't have the code in my head but I think I used a \"multiplier\" of 1.8 - i.e. difference of 1 class would be penalized with 1.8, 2 classes 1.8^2 etc.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 582484,
      "author_name": "Filemon",
      "author_url": "",
      "post_date": "2019-07-23T08:26:38.133000",
      "content": "<p>0.739 - ResNet 50 - official train data - classification - single fold.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 582415,
      "author_name": "Drei",
      "author_url": "",
      "post_date": "2019-07-23T06:46:15.733000",
      "content": "<p>Not doing especially great - .689 using ResNext, external data, kappa optimization, but no TTA or CV (classification only). Some pointers would be greatly appreciated.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 582416,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-07-23T06:46:57.713000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 582069,
      "author_name": "ilovescience",
      "author_url": "",
      "post_date": "2019-07-22T18:08:25.833000",
      "content": "<p>My current score (0.786, 5fold, with external) is classification only</p>",
      "votes": 1,
      "replies": [
        {
          "id": 582705,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-07-23T13:41:19.787000",
          "content": "<p>Thx, how much boost you got with external. I did only one trial and it added 0.025 to my setup at the time. (.774---&gt;.799</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 582989,
          "author_name": "ilovescience",
          "author_url": "",
          "post_date": "2019-07-23T21:21:15.510000",
          "content": "<p>I think I got from 0.705 to 0.760, single fold</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 583222,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-07-24T07:16:33.883000",
          "content": "<p>That's encouraging.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 585266,
      "author_name": "Gary",
      "author_url": "",
      "post_date": "2019-07-27T07:28:53.820000",
      "content": "<p>Hi，did you use the ben's preprocessing?</p>",
      "votes": -1,
      "replies": [
        {
          "id": 585368,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-07-27T10:22:28.737000",
          "content": "<p>I did not</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 585449,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2019-07-27T13:34:25.907000",
          "content": "<p>yep, thank you, but I think the deep model is easy to overfitting for this competition.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 617106,
      "author_name": "Basit Riaz Sheikh",
      "author_url": "",
      "post_date": "2019-09-03T18:41:38.810000",
      "content": "<p><a href=\"/valanm\">@valanm</a> excellent single-model classification results with competition data alone. Do you mind sharing which model architecture you are using? Efficient nets or other? Also, your ensemble seems to get you 0.83+, which 3 models are you using for that? Are these all trained on competition data only?\nI am getting 0.77 with efficient b3 with image size 300 using this competition data only. Trying to hit at least 0.8 with competition data, but struggling...</p>",
      "votes": 0,
      "replies": [
        {
          "id": 622871,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-09-10T07:23:29.733000",
          "content": "<p>Please read my solution</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 616377,
      "author_name": "Miroslav Valan",
      "author_url": "",
      "post_date": "2019-09-03T04:37:17.533000",
      "content": "<p>Update: 0.831 average of 3 classification models. </p>",
      "votes": 0,
      "replies": [
        {
          "id": 617094,
          "author_name": "Basit Riaz Sheikh",
          "author_url": "",
          "post_date": "2019-09-03T18:35:19.707000",
          "content": "<p>You seem to be getting pretty high scores using only the competition data with classification models. Is this ensemble based on competition data only as well, or does this include models pre-trained on old competition data?\nAre using TTA?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 612788,
      "author_name": "GwinHou",
      "author_url": "",
      "post_date": "2019-08-30T02:20:08.323000",
      "content": "<p>Hi, do you guys change to multi class label? \nexp:  label 4 == [1 1 1 1 0]\n         label 5 == [1 1 1 1 1]\nI get a bad score 0.726, by using ben's crop, multi lables, classification, kappa score.  </p>",
      "votes": 0,
      "replies": [
        {
          "id": 612911,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-08-30T04:49:48.917000",
          "content": "<p>No</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 612974,
          "author_name": "GwinHou",
          "author_url": "",
          "post_date": "2019-08-30T06:06:58.800000",
          "content": "<p>Thanks 😬 </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 606053,
      "author_name": "Miroslav Valan",
      "author_url": "",
      "post_date": "2019-08-23T06:13:31.517000",
      "content": "<p>0.817 official data only; simple classification</p>",
      "votes": 0,
      "replies": [
        {
          "id": 606069,
          "author_name": "JIANJIAN",
          "author_url": "",
          "post_date": "2019-08-23T06:40:51.577000",
          "content": "<p>May I ask what is your highest accuracy rate base on CV?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 606214,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-08-23T10:27:14.247000",
          "content": "<p>0.925+ kappa 0.86+ accuracy</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 606223,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2019-08-23T10:45:39.247000",
          "content": "<p>Hi, what model did you use? and 5 fold?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 584369,
      "author_name": "Tahsin Mostafiz",
      "author_url": "",
      "post_date": "2019-07-25T20:38:29.533000",
      "content": "<p>This may sound a little silly but are you guys treating this as a plain old classification task with one hot vector outputs or as a multilabel classification task like <a href=\"https://www.kaggle.com/lextoumbourou/blindness-detection-resnet34-ordinal-targets\">this</a> one?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 584373,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-07-25T20:44:14.627000",
          "content": "<p>plain old classification task in my case :) until  this moment </p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 583677,
      "author_name": "ilovescience",
      "author_url": "",
      "post_date": "2019-07-24T19:56:03.133000",
      "content": "<p>What architecture are you using?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 583700,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-07-24T21:03:42.827000",
          "content": "<p>resnext101</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 583707,
          "author_name": "ilovescience",
          "author_url": "",
          "post_date": "2019-07-24T21:20:11.893000",
          "content": "<p>Ok are you using a lot of different tricks, or more or less out-of-the-box? </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 584379,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-07-25T20:59:47.210000",
          "content": "<p>i do use some minor tricks compared to default settings but nothing that can't be seen in public kernels</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 584410,
          "author_name": "ilovescience",
          "author_url": "",
          "post_date": "2019-07-25T22:26:23.397000",
          "content": "<p>Nice! Thanks!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 583601,
      "author_name": "ABoyko",
      "author_url": "",
      "post_date": "2019-07-24T17:26:06.910000",
      "content": "<p><code>0.795 official data only; classification</code> </p>\n\n<p>Thank you for sharing the information, you can learn more.\nDid you get this result using the classification into 5 classes? Without using regression?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 583602,
          "author_name": "Miroslav Valan",
          "author_url": "",
          "post_date": "2019-07-24T17:29:54.530000",
          "content": "<p>Yes. Only classification trained on official data</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 583608,
          "author_name": "ABoyko",
          "author_url": "",
          "post_date": "2019-07-24T17:41:41.450000",
          "content": "<p>Thanks for the answer. I hope you make your result even better, good luck</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "581851": "There is another topic where teams share their scores and CV vs LB correlations. It seems that most of the top teams use regression. Thus, I am creating this separate topic for those who did some experiments with classification and are willing to share their experience using (1) official data, (2) external data and (3) correlation with regression.\n\n1. 0.789 official train set (5fold). - 0.778 single fold\n2. 0.799 with external (5fold after single trial)\n3. haven't tried yet - will give it a shot soon\n\nUPDATE 2019-07-23:\n1. 0.795 official data only; classification\n\nUPDATE 2019-08-23:\n1. 0.817 official data only; simple classification\n",
    "616386": "I'm also doing only classification, so .834 is possible with an ensemble :)\n(pre-trained on 2015, fine tune on 2019, in total 4 model weights - 2x B4, 1x B3, 1x resnext, TTA, no pseudo labelling yet)",
    "582484": "0.739 - ResNet 50 - official train data - classification - single fold.",
    "582415": "Not doing especially great - .689 using ResNext, external data, kappa optimization, but no TTA or CV (classification only). Some pointers would be greatly appreciated.",
    "582069": "My current score (0.786, 5fold, with external) is classification only",
    "585266": "Hi，did you use the ben's preprocessing?",
    "617106": "@valanm excellent single-model classification results with competition data alone. Do you mind sharing which model architecture you are using? Efficient nets or other? Also, your ensemble seems to get you 0.83+, which 3 models are you using for that? Are these all trained on competition data only?\nI am getting 0.77 with efficient b3 with image size 300 using this competition data only. Trying to hit at least 0.8 with competition data, but struggling...",
    "616377": "Update: 0.831 average of 3 classification models. \n\n",
    "612788": "Hi, do you guys change to multi class label? \nexp:  label 4 == [1 1 1 1 0]\n         label 5 == [1 1 1 1 1]\nI get a bad score 0.726, by using ben's crop, multi lables, classification, kappa score.  ",
    "606053": " 0.817 official data only; simple classification",
    "584369": "This may sound a little silly but are you guys treating this as a plain old classification task with one hot vector outputs or as a multilabel classification task like [this](https://www.kaggle.com/lextoumbourou/blindness-detection-resnet34-ordinal-targets) one?",
    "583677": "What architecture are you using?",
    "583601": "`0.795 official data only; classification` \n\nThank you for sharing the information, you can learn more.\nDid you get this result using the classification into 5 classes? Without using regression?"
  }
}