{
  "id": 170364,
  "title": "Notebook to use Images and Metadata in single EfNet B6 Model",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/170364",
  "author_name": "",
  "post_date": "2020-07-27T12:11:11.237875100Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi all,</p>\n\n<p>I read in the discussion <a href=\"https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/169728\">here </a> from @cdeotte that we can use metadata along with images and can get better results. So, I thought of creating a kernel which does the same. You can add more features as well in the metadata. I only used three features only. </p>\n\n<p>Most of the kernel is stolen from @cdeotte 's original kernel. I changed the model and data pipeline in order to use both. Please read the kernel <a href=\"https://www.kaggle.com/urvishp80/images-and-metadata-in-single-efnet-b6/notebook\">here</a></p>\n\n<p>Upvote it if you like it. </p>",
  "messages": [
    {
      "id": "947659",
      "postDate": "07/27/2020 12:11:11",
      "content": "<p>Hi all,</p>\n\n<p>I read in the discussion <a href=\"https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/169728\">here </a> from @cdeotte that we can use metadata along with images and can get better results. So, I thought of creating a kernel which does the same. You can add more features as well in the metadata. I only used three features only. </p>\n\n<p>Most of the kernel is stolen from @cdeotte 's original kernel. I changed the model and data pipeline in order to use both. Please read the kernel <a href=\"https://www.kaggle.com/urvishp80/images-and-metadata-in-single-efnet-b6/notebook\">here</a></p>\n\n<p>Upvote it if you like it. </p>",
      "rawMarkdown": "Hi all,\n\nI read in the discussion [here ](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/169728) from @cdeotte that we can use metadata along with images and can get better results. So, I thought of creating a kernel which does the same. You can add more features as well in the metadata. I only used three features only. \n\nMost of the kernel is stolen from @cdeotte 's original kernel. I changed the model and data pipeline in order to use both. Please read the kernel [here](https://www.kaggle.com/urvishp80/images-and-metadata-in-single-efnet-b6/notebook)\n\nUpvote it if you like it.",
      "votes": null
    },
    {
      "id": "947668",
      "postDate": "07/27/2020 12:18:33",
      "content": "<p>Whoa you used mixed precision on TPU. Nice work.</p>",
      "rawMarkdown": "Whoa you used mixed precision on TPU. Nice work.",
      "votes": null
    },
    {
      "id": "947672",
      "postDate": "07/27/2020 12:20:10",
      "content": "<p>I did not. I wanted to but then left that part. If you see then it is never enabled. But, you can do it. The only thing is that you need to convert all of data to <code>tf.bfloat16</code> which is not hard. You can use <code>tf.cast</code> for that. I converted metadata into float32 to match with the images. But, you can convert all into bfloat16. </p>",
      "rawMarkdown": "I did not. I wanted to but then left that part. If you see then it is never enabled. But, you can do it. The only thing is that you need to convert all of data to `tf.bfloat16` which is not hard. You can use `tf.cast` for that. I converted metadata into float32 to match with the images. But, you can convert all into bfloat16.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 947668,
      "author_name": "hiramcho",
      "author_url": "",
      "post_date": "07/27/2020 12:18:33",
      "content": "<p>Whoa you used mixed precision on TPU. Nice work.</p>",
      "votes": null,
      "replies": [
        {
          "id": 947672,
          "author_name": "urvishp80",
          "author_url": "",
          "post_date": "07/27/2020 12:20:10",
          "content": "<p>I did not. I wanted to but then left that part. If you see then it is never enabled. But, you can do it. The only thing is that you need to convert all of data to <code>tf.bfloat16</code> which is not hard. You can use <code>tf.cast</code> for that. I converted metadata into float32 to match with the images. But, you can convert all into bfloat16. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "947659": "Hi all,\n\nI read in the discussion [here ](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/169728) from @cdeotte that we can use metadata along with images and can get better results. So, I thought of creating a kernel which does the same. You can add more features as well in the metadata. I only used three features only. \n\nMost of the kernel is stolen from @cdeotte 's original kernel. I changed the model and data pipeline in order to use both. Please read the kernel [here](https://www.kaggle.com/urvishp80/images-and-metadata-in-single-efnet-b6/notebook)\n\nUpvote it if you like it.",
    "947668": "Whoa you used mixed precision on TPU. Nice work.",
    "947672": "I did not. I wanted to but then left that part. If you see then it is never enabled. But, you can do it. The only thing is that you need to convert all of data to `tf.bfloat16` which is not hard. You can use `tf.cast` for that. I converted metadata into float32 to match with the images. But, you can convert all into bfloat16."
  },
  "source": "meta"
}