{
  "id": 173139,
  "title": "Multiple-input models (again?)",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/173139",
  "author_name": "Trigram",
  "post_date": "2020-08-08T03:15:13.749000",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>As we have seen in the recent SIIM-ISIC Melanoma Classification competition, multiple-input models are being used to great effect. So how feasible exactly would it be to use a multi-input model here (i.e. to handle both the confidence and the FVC)</p>\n<p>Here are some ideas I have on how to do this:<br>\nA. Use TFRecords and combine metadata <br>\nB. Use a multi-input CNN as described in Nitesh Chaudhary's notebook.</p>",
  "messages": [
    {
      "id": 962307,
      "postDate": "2020-08-08T03:15:13.750Z",
      "content": "<p>As we have seen in the recent SIIM-ISIC Melanoma Classification competition, multiple-input models are being used to great effect. So how feasible exactly would it be to use a multi-input model here (i.e. to handle both the confidence and the FVC)</p>\n<p>Here are some ideas I have on how to do this:<br>\nA. Use TFRecords and combine metadata <br>\nB. Use a multi-input CNN as described in Nitesh Chaudhary's notebook.</p>",
      "rawMarkdown": "As we have seen in the recent SIIM-ISIC Melanoma Classification competition, multiple-input models are being used to great effect. So how feasible exactly would it be to use a multi-input model here (i.e. to handle both the confidence and the FVC)\n\nHere are some ideas I have on how to do this:\nA. Use TFRecords and combine metadata \nB. Use a multi-input CNN as described in Nitesh Chaudhary's notebook.",
      "votes": 2
    },
    {
      "id": 963433,
      "postDate": "2020-08-09T03:18:16.240Z",
      "content": "<p>What do you mean by multiple-input models?</p>",
      "rawMarkdown": "What do you mean by multiple-input models?",
      "replies": [
        {
          "id": 963722,
          "postDate": "2020-08-09T08:31:46.660Z",
          "content": "<p>Using both tabular and image data with a single model (like in SIIM)</p>",
          "rawMarkdown": "Using both tabular and image data with a single model (like in SIIM)",
          "votes": 1
        }
      ]
    },
    {
      "id": 964161,
      "postDate": "2020-08-09T16:36:45.283Z",
      "content": "<p>Mutliple input models are pretty common when you have multiple sources of data. Either 2 different images or different types of data. It often gives layers the model can use to 'preprocess' them differently before you inevitably merge them into a single layer at some point in the pipeline. Makes logical sense data that are differently structured often would benefit when the gradients in the training are more focused on variables that are more alike. </p>\n\n<p>Not very formal, but that's the intuition I've built up.</p>",
      "rawMarkdown": "Mutliple input models are pretty common when you have multiple sources of data. Either 2 different images or different types of data. It often gives layers the model can use to 'preprocess' them differently before you inevitably merge them into a single layer at some point in the pipeline. Makes logical sense data that are differently structured often would benefit when the gradients in the training are more focused on variables that are more alike. \n\nNot very formal, but that's the intuition I've built up.",
      "votes": 1,
      "isDeleted": true
    },
    {
      "id": 963791,
      "postDate": "2020-08-09T09:35:22.540Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 963433,
      "author_name": "johnny",
      "author_url": "",
      "post_date": "2020-08-09T03:18:16.240000",
      "content": "<p>What do you mean by multiple-input models?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 963722,
          "author_name": "Trigram",
          "author_url": "",
          "post_date": "2020-08-09T08:31:46.660000",
          "content": "<p>Using both tabular and image data with a single model (like in SIIM)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 964161,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-08-09T16:36:45.283000",
      "content": "<p>Mutliple input models are pretty common when you have multiple sources of data. Either 2 different images or different types of data. It often gives layers the model can use to 'preprocess' them differently before you inevitably merge them into a single layer at some point in the pipeline. Makes logical sense data that are differently structured often would benefit when the gradients in the training are more focused on variables that are more alike. </p>\n\n<p>Not very formal, but that's the intuition I've built up.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 963791,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-08-09T09:35:22.540000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "962307": "As we have seen in the recent SIIM-ISIC Melanoma Classification competition, multiple-input models are being used to great effect. So how feasible exactly would it be to use a multi-input model here (i.e. to handle both the confidence and the FVC)\n\nHere are some ideas I have on how to do this:\nA. Use TFRecords and combine metadata \nB. Use a multi-input CNN as described in Nitesh Chaudhary's notebook.",
    "963433": "What do you mean by multiple-input models?",
    "964161": "Mutliple input models are pretty common when you have multiple sources of data. Either 2 different images or different types of data. It often gives layers the model can use to 'preprocess' them differently before you inevitably merge them into a single layer at some point in the pipeline. Makes logical sense data that are differently structured often would benefit when the gradients in the training are more focused on variables that are more alike. \n\nNot very formal, but that's the intuition I've built up.",
    "963791": ""
  }
}