{
  "id": 173479,
  "title": "Best single model 0.953 but bad Ensembles, suggestions ? ",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/173479",
  "author_name": "",
  "post_date": "2020-08-09T13:18:02.811243Z",
  "votes": 4,
  "comment_count": 8,
  "views": 0,
  "content": "<p>I am using the triple-stratified data set with 2018 external data with best single model 0.953. I am not an expert in ensembles, I have tried some methods though such as \nsimple averaging\nweighted averaging \nnp log2 ensemble\npower averaging\nI have only noticed slight improvement +0.003 but not that significant. Any suggestions that worked for you ? </p>",
  "messages": [
    {
      "id": "963962",
      "postDate": "08/09/2020 13:18:02",
      "content": "<p>I am using the triple-stratified data set with 2018 external data with best single model 0.953. I am not an expert in ensembles, I have tried some methods though such as \nsimple averaging\nweighted averaging \nnp log2 ensemble\npower averaging\nI have only noticed slight improvement +0.003 but not that significant. Any suggestions that worked for you ? </p>",
      "rawMarkdown": "I am using the triple-stratified data set with 2018 external data with best single model 0.953. I am not an expert in ensembles, I have tried some methods though such as \nsimple averaging\nweighted averaging \nnp log2 ensemble\npower averaging\nI have only noticed slight improvement +0.003 but not that significant. Any suggestions that worked for you ?",
      "votes": null
    },
    {
      "id": "963964",
      "postDate": "08/09/2020 13:24:35",
      "content": "<p>I have used simple averaging that's work fine.offcourse any other methods can work best.</p>",
      "rawMarkdown": "I have used simple averaging that's work fine.offcourse any other methods can work best.",
      "votes": null
    },
    {
      "id": "964001",
      "postDate": "08/09/2020 14:03:26",
      "content": "<p>Which image size did you use???</p>",
      "rawMarkdown": "Which image size did you use???",
      "votes": null
    },
    {
      "id": "964008",
      "postDate": "08/09/2020 14:10:54",
      "content": "<p>Key for ensembling is diversity in models.  If you average similar model predictions then the upside is small.  Maybe that's what you are seeing?</p>",
      "rawMarkdown": "Key for ensembling is diversity in models.  If you average similar model predictions then the upside is small.  Maybe that's what you are seeing?",
      "votes": null
    },
    {
      "id": "964023",
      "postDate": "08/09/2020 14:20:43",
      "content": "<p>For that particular result 512x512.</p>",
      "rawMarkdown": "For that particular result 512x512.",
      "votes": null
    },
    {
      "id": "964029",
      "postDate": "08/09/2020 14:26:26",
      "content": "<p>Thank you! The current ensemble I am using has different trained models but same resolution. I will definitely try to include a variety of resolutions. </p>",
      "rawMarkdown": "Thank you! The current ensemble I am using has different trained models but same resolution. I will definitely try to include a variety of resolutions.",
      "votes": null
    },
    {
      "id": "964182",
      "postDate": "08/09/2020 16:45:49",
      "content": "<p>If you are using pytorch, change the seed, otherwise the batch that your model learn will probably the same for all your models </p>",
      "rawMarkdown": "If you are using pytorch, change the seed, otherwise the batch that your model learn will probably the same for all your models",
      "votes": null
    },
    {
      "id": "964200",
      "postDate": "08/09/2020 16:52:26",
      "content": "<p>can consider ensemble the result from models with different img size, model structure, loss function…etc</p>",
      "rawMarkdown": "can consider ensemble the result from models with different img size, model structure, loss function...etc",
      "votes": null
    },
    {
      "id": "964201",
      "postDate": "08/09/2020 16:53:28",
      "content": "<p><a href=\"https://www.kaggle.com/shivam17818\" target=\"_blank\">@shivam17818</a>  your user name is creative😂</p>",
      "rawMarkdown": "shivam17818  your user name is creative😂",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 963964,
      "author_name": "shivam17818",
      "author_url": "",
      "post_date": "08/09/2020 13:24:35",
      "content": "<p>I have used simple averaging that's work fine.offcourse any other methods can work best.</p>",
      "votes": null,
      "replies": [
        {
          "id": 964201,
          "author_name": "fiyeroleung",
          "author_url": "",
          "post_date": "08/09/2020 16:53:28",
          "content": "<p><a href=\"https://www.kaggle.com/shivam17818\" target=\"_blank\">@shivam17818</a>  your user name is creative😂</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 964008,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "08/09/2020 14:10:54",
      "content": "<p>Key for ensembling is diversity in models.  If you average similar model predictions then the upside is small.  Maybe that's what you are seeing?</p>",
      "votes": null,
      "replies": [
        {
          "id": 964029,
          "author_name": "alhasanabdellatif123",
          "author_url": "",
          "post_date": "08/09/2020 14:26:26",
          "content": "<p>Thank you! The current ensemble I am using has different trained models but same resolution. I will definitely try to include a variety of resolutions. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 964182,
      "author_name": "ludovick",
      "author_url": "",
      "post_date": "08/09/2020 16:45:49",
      "content": "<p>If you are using pytorch, change the seed, otherwise the batch that your model learn will probably the same for all your models </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 964200,
      "author_name": "fiyeroleung",
      "author_url": "",
      "post_date": "08/09/2020 16:52:26",
      "content": "<p>can consider ensemble the result from models with different img size, model structure, loss function…etc</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 964001,
      "author_name": "msharuk589",
      "author_url": "",
      "post_date": "08/09/2020 14:03:26",
      "content": "<p>Which image size did you use???</p>",
      "votes": null,
      "replies": [
        {
          "id": 964023,
          "author_name": "alhasanabdellatif123",
          "author_url": "",
          "post_date": "08/09/2020 14:20:43",
          "content": "<p>For that particular result 512x512.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "963962": "I am using the triple-stratified data set with 2018 external data with best single model 0.953. I am not an expert in ensembles, I have tried some methods though such as \nsimple averaging\nweighted averaging \nnp log2 ensemble\npower averaging\nI have only noticed slight improvement +0.003 but not that significant. Any suggestions that worked for you ?",
    "963964": "I have used simple averaging that's work fine.offcourse any other methods can work best.",
    "964001": "Which image size did you use???",
    "964008": "Key for ensembling is diversity in models.  If you average similar model predictions then the upside is small.  Maybe that's what you are seeing?",
    "964023": "For that particular result 512x512.",
    "964029": "Thank you! The current ensemble I am using has different trained models but same resolution. I will definitely try to include a variety of resolutions.",
    "964182": "If you are using pytorch, change the seed, otherwise the batch that your model learn will probably the same for all your models",
    "964200": "can consider ensemble the result from models with different img size, model structure, loss function...etc",
    "964201": "shivam17818  your user name is creative😂"
  },
  "source": "meta"
}