{
  "id": 21619,
  "title": "numerical  precision of submission cvs file",
  "url": "/competitions/state-farm-distracted-driver-detection/discussion/21619",
  "author_name": "hengck23",
  "post_date": "2016-06-12T11:38:59.887000",
  "votes": 3,
  "comment_count": 0,
  "views": 335,
  "content": "<p>Hi,</p>\n\n<p>Although the sample submission cvs file is up to 0.1 precision, do use larger number of decimal places for submission, e.g. up to 0.000001. </p>\n\n<p>I find that the log loss is quite sensitive to this. I experiment with finetunning googlenet. The model gives log loss core of 2.64 and 0.59 when using submission files of 0.1 and 0.00001 precision respectively.</p>\n\n<p>Hope this helps!</p>",
  "messages": [
    {
      "id": 123511,
      "postDate": "2016-06-12T11:38:59.887Z",
      "content": "<p>Hi,</p>\n\n<p>Although the sample submission cvs file is up to 0.1 precision, do use larger number of decimal places for submission, e.g. up to 0.000001. </p>\n\n<p>I find that the log loss is quite sensitive to this. I experiment with finetunning googlenet. The model gives log loss core of 2.64 and 0.59 when using submission files of 0.1 and 0.00001 precision respectively.</p>\n\n<p>Hope this helps!</p>",
      "rawMarkdown": "Hi,\r\n\r\nAlthough the sample submission cvs file is up to 0.1 precision, do use larger number of decimal places for submission, e.g. up to 0.000001. \r\n\r\nI find that the log loss is quite sensitive to this. I experiment with finetunning googlenet. The model gives log loss core of 2.64 and 0.59 when using submission files of 0.1 and 0.00001 precision respectively.\r\n\r\nHope this helps!",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "123511": "Hi,\r\n\r\nAlthough the sample submission cvs file is up to 0.1 precision, do use larger number of decimal places for submission, e.g. up to 0.000001. \r\n\r\nI find that the log loss is quite sensitive to this. I experiment with finetunning googlenet. The model gives log loss core of 2.64 and 0.59 when using submission files of 0.1 and 0.00001 precision respectively.\r\n\r\nHope this helps!"
  }
}