{
  "id": 170261,
  "title": "RIDGE/LGB/NN/ on meta data Optuna focal loss ",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/170261",
  "author_name": "",
  "post_date": "2020-07-27T04:21:53.590043Z",
  "votes": 9,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n\n<p>I created this kernel for combining ridge/LGB/NN on meta data together with Optuna.\nThis is my first kernel. Please check and give me feedback: \n<a href=\"https://www.kaggle.com/chen2222/ridge-lgb-nn-on-meta-data-optuna-focal-loss\">https://www.kaggle.com/chen2222/ridge-lgb-nn-on-meta-data-optuna-focal-loss</a></p>",
  "messages": [
    {
      "id": "947082",
      "postDate": "07/27/2020 04:21:53",
      "content": "<p>Hi everyone,</p>\n\n<p>I created this kernel for combining ridge/LGB/NN on meta data together with Optuna.\nThis is my first kernel. Please check and give me feedback: \n<a href=\"https://www.kaggle.com/chen2222/ridge-lgb-nn-on-meta-data-optuna-focal-loss\">https://www.kaggle.com/chen2222/ridge-lgb-nn-on-meta-data-optuna-focal-loss</a></p>",
      "rawMarkdown": "Hi everyone,\n\nI created this kernel for combining ridge/LGB/NN on meta data together with Optuna.\nThis is my first kernel. Please check and give me feedback: \nhttps://www.kaggle.com/chen2222/ridge-lgb-nn-on-meta-data-optuna-focal-loss",
      "votes": null
    },
    {
      "id": "947085",
      "postDate": "07/27/2020 04:25:21",
      "content": "<p>Cool notebook. It looks like you're ensembling Ridge/LGBM/and MLP using only meta data. What's the CV LB?</p>",
      "rawMarkdown": "Cool notebook. It looks like you're ensembling Ridge/LGBM/and MLP using only meta data. What's the CV LB?",
      "votes": null
    },
    {
      "id": "947109",
      "postDate": "07/27/2020 04:48:44",
      "content": "<p>I didn't explicitly check CV, but I believe should be close to LB. The LB for the ensembled model is 0.6978. </p>",
      "rawMarkdown": "I didn't explicitly check CV, but I believe should be close to LB. The LB for the ensembled model is 0.6978.",
      "votes": null
    },
    {
      "id": "947112",
      "postDate": "07/27/2020 04:52:00",
      "content": "<p><a href=\"/chen2222\">@chen2222</a> Was <strong>optuna</strong> faster and gave better hyper parameters compared to Grid Search? </p>",
      "rawMarkdown": "chen2222 Was **optuna** faster and gave better hyper parameters compared to Grid Search?",
      "votes": null
    },
    {
      "id": "947139",
      "postDate": "07/27/2020 05:06:07",
      "content": "<p>I gave it a go and got much the same, LB 0.6970</p>",
      "rawMarkdown": "I gave it a go and got much the same, LB 0.6970",
      "votes": null
    },
    {
      "id": "947159",
      "postDate": "07/27/2020 05:25:01",
      "content": "<p>Yes, it is faster than gridsearch. </p>",
      "rawMarkdown": "Yes, it is faster than gridsearch.",
      "votes": null
    },
    {
      "id": "947162",
      "postDate": "07/27/2020 05:30:02",
      "content": "<p>Optuna uses Tpesampler which was the one I used. It will find the best parameters faster. You can check their web for more details. Please give my notebook a vote :s</p>",
      "rawMarkdown": "Optuna uses Tpesampler which was the one I used. It will find the best parameters faster. You can check their web for more details. Please give my notebook a vote :s",
      "votes": null
    },
    {
      "id": "948476",
      "postDate": "07/28/2020 02:00:15",
      "content": "<p><a href=\"/mutantspore\">@mutantspore</a> Instead of using model trained on meta directly. You could also take this stacking approach  <a href=\"https://www.kaggle.com/chen2222/lgb-mlp-ridge-stacking-effnet-oof-with-meta\">https://www.kaggle.com/chen2222/lgb-mlp-ridge-stacking-effnet-oof-with-meta</a></p>",
      "rawMarkdown": "mutantspore Instead of using model trained on meta directly. You could also take this stacking approach  https://www.kaggle.com/chen2222/lgb-mlp-ridge-stacking-effnet-oof-with-meta",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 947085,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "07/27/2020 04:25:21",
      "content": "<p>Cool notebook. It looks like you're ensembling Ridge/LGBM/and MLP using only meta data. What's the CV LB?</p>",
      "votes": null,
      "replies": [
        {
          "id": 947109,
          "author_name": "chen2222",
          "author_url": "",
          "post_date": "07/27/2020 04:48:44",
          "content": "<p>I didn't explicitly check CV, but I believe should be close to LB. The LB for the ensembled model is 0.6978. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 947139,
          "author_name": "mutantspore",
          "author_url": "",
          "post_date": "07/27/2020 05:06:07",
          "content": "<p>I gave it a go and got much the same, LB 0.6970</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 948476,
          "author_name": "chen2222",
          "author_url": "",
          "post_date": "07/28/2020 02:00:15",
          "content": "<p><a href=\"/mutantspore\">@mutantspore</a> Instead of using model trained on meta directly. You could also take this stacking approach  <a href=\"https://www.kaggle.com/chen2222/lgb-mlp-ridge-stacking-effnet-oof-with-meta\">https://www.kaggle.com/chen2222/lgb-mlp-ridge-stacking-effnet-oof-with-meta</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 947112,
      "author_name": "sirishks",
      "author_url": "",
      "post_date": "07/27/2020 04:52:00",
      "content": "<p><a href=\"/chen2222\">@chen2222</a> Was <strong>optuna</strong> faster and gave better hyper parameters compared to Grid Search? </p>",
      "votes": null,
      "replies": [
        {
          "id": 947159,
          "author_name": "chen2222",
          "author_url": "",
          "post_date": "07/27/2020 05:25:01",
          "content": "<p>Yes, it is faster than gridsearch. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 947162,
          "author_name": "chen2222",
          "author_url": "",
          "post_date": "07/27/2020 05:30:02",
          "content": "<p>Optuna uses Tpesampler which was the one I used. It will find the best parameters faster. You can check their web for more details. Please give my notebook a vote :s</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "947082": "Hi everyone,\n\nI created this kernel for combining ridge/LGB/NN on meta data together with Optuna.\nThis is my first kernel. Please check and give me feedback: \nhttps://www.kaggle.com/chen2222/ridge-lgb-nn-on-meta-data-optuna-focal-loss",
    "947085": "Cool notebook. It looks like you're ensembling Ridge/LGBM/and MLP using only meta data. What's the CV LB?",
    "947109": "I didn't explicitly check CV, but I believe should be close to LB. The LB for the ensembled model is 0.6978.",
    "947112": "chen2222 Was **optuna** faster and gave better hyper parameters compared to Grid Search?",
    "947139": "I gave it a go and got much the same, LB 0.6970",
    "947159": "Yes, it is faster than gridsearch.",
    "947162": "Optuna uses Tpesampler which was the one I used. It will find the best parameters faster. You can check their web for more details. Please give my notebook a vote :s",
    "948476": "mutantspore Instead of using model trained on meta directly. You could also take this stacking approach  https://www.kaggle.com/chen2222/lgb-mlp-ridge-stacking-effnet-oof-with-meta"
  },
  "source": "meta"
}