{
  "id": 167945,
  "title": "Strategy for Meta-Data to increase LB score 0.001 ",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/167945",
  "author_name": "Mohamad Merchant",
  "post_date": "2020-07-18T12:54:03.582000",
  "votes": 8,
  "comment_count": 11,
  "views": 0,
  "content": "<ol>\n<li>Try simple feature engineering.</li>\n<li>Use XGB, LGBM, Catboost, TabNet.</li>\n<li>Random search hyperparams and train all models on 5 folds CV.</li>\n<li>Make an ensemble of all models with some post processing. (Try different techniques).</li>\n<li>Weighted Average with your images submission. (Try with weights 0.1 to 0.3 on meta-data)</li>\n</ol>",
  "messages": [
    {
      "id": 934445,
      "postDate": "2020-07-18T12:54:03.583Z",
      "content": "<ol>\n<li>Try simple feature engineering.</li>\n<li>Use XGB, LGBM, Catboost, TabNet.</li>\n<li>Random search hyperparams and train all models on 5 folds CV.</li>\n<li>Make an ensemble of all models with some post processing. (Try different techniques).</li>\n<li>Weighted Average with your images submission. (Try with weights 0.1 to 0.3 on meta-data)</li>\n</ol>",
      "rawMarkdown": "1. Try simple feature engineering.\n2. Use XGB, LGBM, Catboost, TabNet.\n3. Random search hyperparams and train all models on 5 folds CV.\n4. Make an ensemble of all models with some post processing. (Try different techniques).\n5.  Weighted Average with your images submission. (Try with weights 0.1 to 0.3 on meta-data)\n",
      "votes": 7
    },
    {
      "id": 934516,
      "postDate": "2020-07-18T13:59:35.483Z",
      "content": "<p>Please explain what technique you have used to improve accuracy?? </p>",
      "rawMarkdown": "Please explain what technique you have used to improve accuracy?? ",
      "votes": 1
    },
    {
      "id": 934659,
      "postDate": "2020-07-18T16:03:04.543Z",
      "content": "<p>Ensemble is the key. Use different models for images with different sizes and also for meta data.\nAlso try post processing such as mean average, rank average, weighted average on all the submissions.</p>",
      "rawMarkdown": "Ensemble is the key. Use different models for images with different sizes and also for meta data.\nAlso try post processing such as mean average, rank average, weighted average on all the submissions.",
      "votes": 2,
      "replies": [
        {
          "id": 934718,
          "postDate": "2020-07-18T17:05:49.820Z",
          "content": "<p>thank you!</p>",
          "rawMarkdown": "thank you!"
        },
        {
          "id": 939287,
          "postDate": "2020-07-22T06:25:36.787Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 939396,
          "postDate": "2020-07-22T08:06:28.207Z",
          "content": "<p>512x512 on B4 gave me the best results so far.</p>",
          "rawMarkdown": "512x512 on B4 gave me the best results so far.\n\n"
        },
        {
          "id": 939403,
          "postDate": "2020-07-22T08:13:34.917Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 939405,
          "postDate": "2020-07-22T08:15:18.143Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 939571,
          "postDate": "2020-07-22T10:01:10.010Z",
          "content": "<p>0.949\nwith meta features +0.001 i.e 0.95</p>",
          "rawMarkdown": "0.949\nwith meta features +0.001 i.e 0.95"
        },
        {
          "id": 940221,
          "postDate": "2020-07-22T18:41:15.347Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 940577,
          "postDate": "2020-07-23T03:14:35.913Z",
          "content": "<p>hi <a href=\"/mohamadmerchant\">@mohamadmerchant</a> , thanks for this.\nhow much tta do you use?\nin my exp,  tta and lb are not positively correlated. Large tta improves local cv, but lb drops</p>",
          "rawMarkdown": "hi @mohamadmerchant , thanks for this.\nhow much tta do you use?\nin my exp,  tta and lb are not positively correlated. Large tta improves local cv, but lb drops"
        },
        {
          "id": 941243,
          "postDate": "2020-07-23T06:33:34.047Z",
          "content": "<p>Yes, large TTA drops LB i.e TTAx20 didn't performed well for me.\nI have used TTAx15 and TTAx11 both are good for this competition.</p>",
          "rawMarkdown": "Yes, large TTA drops LB i.e TTAx20 didn't performed well for me.\nI have used TTAx15 and TTAx11 both are good for this competition."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 934516,
      "author_name": "Shivam",
      "author_url": "",
      "post_date": "2020-07-18T13:59:35.483000",
      "content": "<p>Please explain what technique you have used to improve accuracy?? </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 934659,
      "author_name": "Mohamad Merchant",
      "author_url": "",
      "post_date": "2020-07-18T16:03:04.543000",
      "content": "<p>Ensemble is the key. Use different models for images with different sizes and also for meta data.\nAlso try post processing such as mean average, rank average, weighted average on all the submissions.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 934718,
          "author_name": "Shivam",
          "author_url": "",
          "post_date": "2020-07-18T17:05:49.820000",
          "content": "<p>thank you!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 939287,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-07-22T06:25:36.787000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 939396,
          "author_name": "Mohamad Merchant",
          "author_url": "",
          "post_date": "2020-07-22T08:06:28.207000",
          "content": "<p>512x512 on B4 gave me the best results so far.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 939403,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-07-22T08:13:34.917000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 939405,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-07-22T08:15:18.143000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 939571,
          "author_name": "Mohamad Merchant",
          "author_url": "",
          "post_date": "2020-07-22T10:01:10.010000",
          "content": "<p>0.949\nwith meta features +0.001 i.e 0.95</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 940221,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-07-22T18:41:15.347000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 940577,
          "author_name": "yunxi",
          "author_url": "",
          "post_date": "2020-07-23T03:14:35.913000",
          "content": "<p>hi <a href=\"/mohamadmerchant\">@mohamadmerchant</a> , thanks for this.\nhow much tta do you use?\nin my exp,  tta and lb are not positively correlated. Large tta improves local cv, but lb drops</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 941243,
          "author_name": "Mohamad Merchant",
          "author_url": "",
          "post_date": "2020-07-23T06:33:34.047000",
          "content": "<p>Yes, large TTA drops LB i.e TTAx20 didn't performed well for me.\nI have used TTAx15 and TTAx11 both are good for this competition.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "934445": "1. Try simple feature engineering.\n2. Use XGB, LGBM, Catboost, TabNet.\n3. Random search hyperparams and train all models on 5 folds CV.\n4. Make an ensemble of all models with some post processing. (Try different techniques).\n5.  Weighted Average with your images submission. (Try with weights 0.1 to 0.3 on meta-data)\n",
    "934516": "Please explain what technique you have used to improve accuracy?? ",
    "934659": "Ensemble is the key. Use different models for images with different sizes and also for meta data.\nAlso try post processing such as mean average, rank average, weighted average on all the submissions."
  }
}