{
  "id": 164354,
  "title": "Are you guys using any external data?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/164354",
  "author_name": "",
  "post_date": "2020-07-06T00:25:34.963564900Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "",
  "messages": [
    {
      "id": "916710",
      "postDate": "07/06/2020 00:25:34",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "916725",
      "postDate": "07/06/2020 01:05:09",
      "content": "<p>I'm using an external dataset given <a href=\"https://www.kaggle.com/shonenkov/merge-external-data/\">here</a> and I believe many people are using it. It gives me a boost of 4-5% on a single efficient net model + augmentation + undersampled data.\nMy LB is 90.025</p>",
      "rawMarkdown": "I'm using an external dataset given [here](https://www.kaggle.com/shonenkov/merge-external-data/) and I believe many people are using it. It gives me a boost of 4-5% on a single efficient net model + augmentation + undersampled data.\nMy LB is 90.025",
      "votes": null
    },
    {
      "id": "916873",
      "postDate": "07/06/2020 04:58:57",
      "content": "<p>Yes, You can add external data but you might need to do feature engineering in order to get some significant boost in the performance. Otherwise, I will focus more on hypertuning the model. You can try to hypertune the effiicient B5 model.</p>",
      "rawMarkdown": "Yes, You can add external data but you might need to do feature engineering in order to get some significant boost in the performance. Otherwise, I will focus more on hypertuning the model. You can try to hypertune the effiicient B5 model.",
      "votes": null
    },
    {
      "id": "917318",
      "postDate": "07/06/2020 12:09:11",
      "content": "<p>But to me it works vise versa. I added an external data and it got me a lower LB (decreased from 93.5 to 92.2).\nI tested that for both 256 and 512 image sizes. I suppose its better to focus on methods (i.e hyper tuning and setting the model) we are using.  </p>",
      "rawMarkdown": "But to me it works vise versa. I added an external data and it got me a lower LB (decreased from 93.5 to 92.2).\nI tested that for both 256 and 512 image sizes. I suppose its better to focus on methods (i.e hyper tuning and setting the model) we are using.",
      "votes": null
    },
    {
      "id": "917347",
      "postDate": "07/06/2020 12:32:12",
      "content": "<p>B5 is a good model to work with, do you have any intuition why?</p>",
      "rawMarkdown": "B5 is a good model to work with, do you have any intuition why?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 916725,
      "author_name": "prateek0x",
      "author_url": "",
      "post_date": "07/06/2020 01:05:09",
      "content": "<p>I'm using an external dataset given <a href=\"https://www.kaggle.com/shonenkov/merge-external-data/\">here</a> and I believe many people are using it. It gives me a boost of 4-5% on a single efficient net model + augmentation + undersampled data.\nMy LB is 90.025</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 916873,
      "author_name": "",
      "author_url": "",
      "post_date": "07/06/2020 04:58:57",
      "content": "<p>Yes, You can add external data but you might need to do feature engineering in order to get some significant boost in the performance. Otherwise, I will focus more on hypertuning the model. You can try to hypertune the effiicient B5 model.</p>",
      "votes": null,
      "replies": [
        {
          "id": 917347,
          "author_name": "treadon",
          "author_url": "",
          "post_date": "07/06/2020 12:32:12",
          "content": "<p>B5 is a good model to work with, do you have any intuition why?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 917318,
      "author_name": "elenadehnavi",
      "author_url": "",
      "post_date": "07/06/2020 12:09:11",
      "content": "<p>But to me it works vise versa. I added an external data and it got me a lower LB (decreased from 93.5 to 92.2).\nI tested that for both 256 and 512 image sizes. I suppose its better to focus on methods (i.e hyper tuning and setting the model) we are using.  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "916710": "",
    "916725": "I'm using an external dataset given [here](https://www.kaggle.com/shonenkov/merge-external-data/) and I believe many people are using it. It gives me a boost of 4-5% on a single efficient net model + augmentation + undersampled data.\nMy LB is 90.025",
    "916873": "Yes, You can add external data but you might need to do feature engineering in order to get some significant boost in the performance. Otherwise, I will focus more on hypertuning the model. You can try to hypertune the effiicient B5 model.",
    "917318": "But to me it works vise versa. I added an external data and it got me a lower LB (decreased from 93.5 to 92.2).\nI tested that for both 256 and 512 image sizes. I suppose its better to focus on methods (i.e hyper tuning and setting the model) we are using.",
    "917347": "B5 is a good model to work with, do you have any intuition why?"
  },
  "source": "meta"
}