{
  "id": 413428,
  "title": "How about Label Smoothing?",
  "url": "/competitions/google-research-identify-contrails-reduce-global-warming/discussion/413428",
  "author_name": "LUPIN11",
  "post_date": "2023-05-28T15:37:54.880000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I wonder whether we can make use of labels in <strong>human_individual_masks.npy</strong>？<br>\nFor Example:</p>\n<pre><code>label = np.load(\"human_individual_masks.npy\")\nk = label.shape[-1]\nlabel = np.sum(label, axis=-1)\nlabel = label / k\n</code></pre>\n<p>This gives us:<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9893459%2Fd75b4a1516537e82583b58e4faf52a09%2Fm.png?generation=1685288157277973&amp;alt=media\" alt=\"\"><br>\nCan such labels provide <strong>better guidance</strong> for the model?<br>\nMore results can be found here: <a href=\"https://www.kaggle.com/code/lupin11/labelsmoothing/notebook\" target=\"_blank\">https://www.kaggle.com/code/lupin11/labelsmoothing/notebook</a></p>",
  "messages": [
    {
      "id": 2278338,
      "postDate": "2023-05-28T15:37:54.880Z",
      "content": "<p>I wonder whether we can make use of labels in <strong>human_individual_masks.npy</strong>？<br>\nFor Example:</p>\n<pre><code>label = np.load(\"human_individual_masks.npy\")\nk = label.shape[-1]\nlabel = np.sum(label, axis=-1)\nlabel = label / k\n</code></pre>\n<p>This gives us:<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9893459%2Fd75b4a1516537e82583b58e4faf52a09%2Fm.png?generation=1685288157277973&amp;alt=media\" alt=\"\"><br>\nCan such labels provide <strong>better guidance</strong> for the model?<br>\nMore results can be found here: <a href=\"https://www.kaggle.com/code/lupin11/labelsmoothing/notebook\" target=\"_blank\">https://www.kaggle.com/code/lupin11/labelsmoothing/notebook</a></p>",
      "rawMarkdown": "I wonder whether we can make use of labels in **human_individual_masks.npy**？\nFor Example:\n```\nlabel = np.load(\"human_individual_masks.npy\")\nk = label.shape[-1]\nlabel = np.sum(label, axis=-1)\nlabel = label / k\n```\nThis gives us:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9893459%2Fd75b4a1516537e82583b58e4faf52a09%2Fm.png?generation=1685288157277973&alt=media)\nCan such labels provide **better guidance** for the model?\nMore results can be found here: [https://www.kaggle.com/code/lupin11/labelsmoothing/notebook](https://www.kaggle.com/code/lupin11/labelsmoothing/notebook)",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2278338": "I wonder whether we can make use of labels in **human_individual_masks.npy**？\nFor Example:\n```\nlabel = np.load(\"human_individual_masks.npy\")\nk = label.shape[-1]\nlabel = np.sum(label, axis=-1)\nlabel = label / k\n```\nThis gives us:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9893459%2Fd75b4a1516537e82583b58e4faf52a09%2Fm.png?generation=1685288157277973&alt=media)\nCan such labels provide **better guidance** for the model?\nMore results can be found here: [https://www.kaggle.com/code/lupin11/labelsmoothing/notebook](https://www.kaggle.com/code/lupin11/labelsmoothing/notebook)"
  }
}