{
  "id": 279031,
  "title": "Lovasz loss, probability calibration and threshold",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/279031",
  "author_name": "mahdi",
  "post_date": "2021-10-16T13:38:57.126000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Technically, Lovasz loss is using hinge loss. Hence the output from the last layer (connecting to the Lovas loss) is not outputting logit value. It is not correct to say prob = sigmoid(output).</p>\n<p>you should do a probability calibration if you need probability output. you can google for \"calibration for SVM output\" for more you want to get prob output.</p>\n<p>this may or may not have any importance at all, depending on how you ensemble and thresholding.</p>\n<hr>\n<p>note:</p>\n<p>prob = sigmoid(output) only if the output is connected to binary cross-entropy loss layer.</p>",
  "messages": [
    {
      "id": 1546757,
      "postDate": "2021-10-16T13:38:57.127Z",
      "content": "<p>Technically, Lovasz loss is using hinge loss. Hence the output from the last layer (connecting to the Lovas loss) is not outputting logit value. It is not correct to say prob = sigmoid(output).</p>\n<p>you should do a probability calibration if you need probability output. you can google for \"calibration for SVM output\" for more you want to get prob output.</p>\n<p>this may or may not have any importance at all, depending on how you ensemble and thresholding.</p>\n<hr>\n<p>note:</p>\n<p>prob = sigmoid(output) only if the output is connected to binary cross-entropy loss layer.</p>",
      "rawMarkdown": "Technically, Lovasz loss is using hinge loss. Hence the output from the last layer (connecting to the Lovas loss) is not outputting logit value. It is not correct to say prob = sigmoid(output).\n\nyou should do a probability calibration if you need probability output. you can google for \"calibration for SVM output\" for more you want to get prob output.\n\nthis may or may not have any importance at all, depending on how you ensemble and thresholding.\n\n- - - - - - - - - - - - \nnote:\n\nprob = sigmoid(output) only if the output is connected to binary cross-entropy loss layer.",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1546757": "Technically, Lovasz loss is using hinge loss. Hence the output from the last layer (connecting to the Lovas loss) is not outputting logit value. It is not correct to say prob = sigmoid(output).\n\nyou should do a probability calibration if you need probability output. you can google for \"calibration for SVM output\" for more you want to get prob output.\n\nthis may or may not have any importance at all, depending on how you ensemble and thresholding.\n\n- - - - - - - - - - - - \nnote:\n\nprob = sigmoid(output) only if the output is connected to binary cross-entropy loss layer."
  }
}