{
  "id": 181274,
  "title": "How to combine meta features with Images?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/181274",
  "author_name": "Amritvir Singh",
  "post_date": "2020-09-08T08:36:49.231000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>How do I use meta-features given in csv to train along with images? </p>",
  "messages": [
    {
      "id": 1002578,
      "postDate": "2020-09-08T08:36:49.230Z",
      "content": "<p>How do I use meta-features given in csv to train along with images? </p>",
      "rawMarkdown": "How do I use meta-features given in csv to train along with images? ",
      "votes": 2
    },
    {
      "id": 1059182,
      "postDate": "2020-10-24T18:39:40.170Z",
      "content": "<p><a href=\"https://www.kaggle.com/amritvirsinghx\" target=\"_blank\">@amritvirsinghx</a> <br>\nHope this may help. <a href=\"https://www.pyimagesearch.com/2019/02/04/keras-multiple-inputs-and-mixed-data/\" target=\"_blank\">Multiple Inputs and Mixed Data</a></p>",
      "rawMarkdown": "@amritvirsinghx \nHope this may help. [Multiple Inputs and Mixed Data](https://www.pyimagesearch.com/2019/02/04/keras-multiple-inputs-and-mixed-data/)"
    },
    {
      "id": 1033334,
      "postDate": "2020-09-30T21:20:00.670Z",
      "content": "<p>You can train once a meta-data based model e.g. on XGBoost and once a model based on image-recognition e.g. EfficientNet-B1, you predict on both of them and you can use e.g. the simple average of both:</p>\n<p><code>score = x1 + x2 / 2</code></p>\n<p>with x1 and x2 being the predicitons of both models.</p>\n<p>Personally I used <code>weighted averaging</code> and calculated my scores like this:</p>\n<p><code>score = x1*0.7 + x2*0.3</code></p>\n<p>because I wanted to value the image recognition (x1) higher than the meta-data recognition (x2)</p>\n<p>During the competiton, a lot of people reported that training it seperately worked better for them instead of using the meta-data within the model.</p>",
      "rawMarkdown": "You can train once a meta-data based model e.g. on XGBoost and once a model based on image-recognition e.g. EfficientNet-B1, you predict on both of them and you can use e.g. the simple average of both:\n\n`score = x1 + x2 / 2`\n\nwith x1 and x2 being the predicitons of both models.\n\nPersonally I used `weighted averaging` and calculated my scores like this:\n\n`score = x1*0.7 + x2*0.3`\n\nbecause I wanted to value the image recognition (x1) higher than the meta-data recognition (x2)\n\nDuring the competiton, a lot of people reported that training it seperately worked better for them instead of using the meta-data within the model."
    },
    {
      "id": 1003108,
      "postDate": "2020-09-08T17:01:21.143Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1059182,
      "author_name": "Innat",
      "author_url": "",
      "post_date": "2020-10-24T18:39:40.170000",
      "content": "<p><a href=\"https://www.kaggle.com/amritvirsinghx\" target=\"_blank\">@amritvirsinghx</a> <br>\nHope this may help. <a href=\"https://www.pyimagesearch.com/2019/02/04/keras-multiple-inputs-and-mixed-data/\" target=\"_blank\">Multiple Inputs and Mixed Data</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1033334,
      "author_name": "Ali Abdin",
      "author_url": "",
      "post_date": "2020-09-30T21:20:00.670000",
      "content": "<p>You can train once a meta-data based model e.g. on XGBoost and once a model based on image-recognition e.g. EfficientNet-B1, you predict on both of them and you can use e.g. the simple average of both:</p>\n<p><code>score = x1 + x2 / 2</code></p>\n<p>with x1 and x2 being the predicitons of both models.</p>\n<p>Personally I used <code>weighted averaging</code> and calculated my scores like this:</p>\n<p><code>score = x1*0.7 + x2*0.3</code></p>\n<p>because I wanted to value the image recognition (x1) higher than the meta-data recognition (x2)</p>\n<p>During the competiton, a lot of people reported that training it seperately worked better for them instead of using the meta-data within the model.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1003108,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-09-08T17:01:21.143000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1002578": "How do I use meta-features given in csv to train along with images? ",
    "1059182": "@amritvirsinghx \nHope this may help. [Multiple Inputs and Mixed Data](https://www.pyimagesearch.com/2019/02/04/keras-multiple-inputs-and-mixed-data/)",
    "1033334": "You can train once a meta-data based model e.g. on XGBoost and once a model based on image-recognition e.g. EfficientNet-B1, you predict on both of them and you can use e.g. the simple average of both:\n\n`score = x1 + x2 / 2`\n\nwith x1 and x2 being the predicitons of both models.\n\nPersonally I used `weighted averaging` and calculated my scores like this:\n\n`score = x1*0.7 + x2*0.3`\n\nbecause I wanted to value the image recognition (x1) higher than the meta-data recognition (x2)\n\nDuring the competiton, a lot of people reported that training it seperately worked better for them instead of using the meta-data within the model.",
    "1003108": ""
  }
}