{
  "id": 215282,
  "title": "Understanding what to use False positives data for ??",
  "url": "/competitions/rfcx-species-audio-detection/discussion/215282",
  "author_name": "",
  "post_date": "2021-01-29T09:57:26.257039900Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Does anybody have an idea what the False positive dataset and the recordings should be used for ??</p>",
  "messages": [
    {
      "id": "1175726",
      "postDate": "01/29/2021 09:57:26",
      "content": "<p>Does anybody have an idea what the False positive dataset and the recordings should be used for ??</p>",
      "rawMarkdown": "Does anybody have an idea what the False positive dataset and the recordings should be used for ??",
      "votes": null
    },
    {
      "id": "1176436",
      "postDate": "01/29/2021 16:19:43",
      "content": "<p>Some of the labelled TPs have other classes present in the crop, as do the FPs. The only thing we know for <em>certain</em> about the FPs is that the particular class is not present. In other words, they are unambiguous data samples for one class only. I've experimented with labels that are 0.5 (or so) for all classes and only 1 or 0 for the class in question, depending on whether or not it's an FP or TP, but it didn't improve my validation procedure. </p>\n<p>Someone cleverer than me has probably found a way to use them to their benefit, but I haven't found them useful.</p>",
      "rawMarkdown": "Some of the labelled TPs have other classes present in the crop, as do the FPs. The only thing we know for *certain* about the FPs is that the particular class is not present. In other words, they are unambiguous data samples for one class only. I've experimented with labels that are 0.5 (or so) for all classes and only 1 or 0 for the class in question, depending on whether or not it's an FP or TP, but it didn't improve my validation procedure. \n\nSomeone cleverer than me has probably found a way to use them to their benefit, but I haven't found them useful.",
      "votes": null
    },
    {
      "id": "1176879",
      "postDate": "01/29/2021 21:21:22",
      "content": "<p>Hi,</p>\n<p>You can use them for data augmentation like background noise. <a href=\"https://github.com/iver56/audiomentations\" target=\"_blank\">audiomentations</a> library does that for you.</p>",
      "rawMarkdown": "Hi,\n\nYou can use them for data augmentation like background noise. [audiomentations](https://github.com/iver56/audiomentations) library does that for you.",
      "votes": null
    },
    {
      "id": "1179637",
      "postDate": "01/31/2021 17:37:49",
      "content": "<p>Article and code that use FP:<br>\n<a href=\"https://github.com/Sieve-Analytics/arbimon2-cnn/blob/master/2_train_ResNet50.ipynb\" target=\"_blank\">https://github.com/Sieve-Analytics/arbimon2-cnn/blob/master/2_train_ResNet50.ipynb</a><br>\n<a href=\"https://www.sciencedirect.com/science/article/pii/S1574954120300637\" target=\"_blank\">https://www.sciencedirect.com/science/article/pii/S1574954120300637</a>  </p>",
      "rawMarkdown": "Article and code that use FP:\nhttps://github.com/Sieve-Analytics/arbimon2-cnn/blob/master/2_train_ResNet50.ipynb\nhttps://www.sciencedirect.com/science/article/pii/S1574954120300637",
      "votes": null
    },
    {
      "id": "1183447",
      "postDate": "02/03/2021 02:25:11",
      "content": "<p>yes, the challenge is that FP data only gives Not Present information for particular class, and this is a multi-label task. I<code>m thinking about to play some trick in training loss function, to mask/suppress 23 lables other than the one indicated as 0 in FP data, to let the network only learn from this particular label 0 indicated from FP data. Don</code>t know if it`s feasible.</p>",
      "rawMarkdown": "yes, the challenge is that FP data only gives Not Present information for particular class, and this is a multi-label task. I`m thinking about to play some trick in training loss function, to mask/suppress 23 lables other than the one indicated as 0 in FP data, to let the network only learn from this particular label 0 indicated from FP data. Don`t know if it`s feasible.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1176436,
      "author_name": "bigironsphere",
      "author_url": "",
      "post_date": "01/29/2021 16:19:43",
      "content": "<p>Some of the labelled TPs have other classes present in the crop, as do the FPs. The only thing we know for <em>certain</em> about the FPs is that the particular class is not present. In other words, they are unambiguous data samples for one class only. I've experimented with labels that are 0.5 (or so) for all classes and only 1 or 0 for the class in question, depending on whether or not it's an FP or TP, but it didn't improve my validation procedure. </p>\n<p>Someone cleverer than me has probably found a way to use them to their benefit, but I haven't found them useful.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1183447,
          "author_name": "superchenhao",
          "author_url": "",
          "post_date": "02/03/2021 02:25:11",
          "content": "<p>yes, the challenge is that FP data only gives Not Present information for particular class, and this is a multi-label task. I<code>m thinking about to play some trick in training loss function, to mask/suppress 23 lables other than the one indicated as 0 in FP data, to let the network only learn from this particular label 0 indicated from FP data. Don</code>t know if it`s feasible.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1176879,
      "author_name": "snnclsr",
      "author_url": "",
      "post_date": "01/29/2021 21:21:22",
      "content": "<p>Hi,</p>\n<p>You can use them for data augmentation like background noise. <a href=\"https://github.com/iver56/audiomentations\" target=\"_blank\">audiomentations</a> library does that for you.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1179637,
      "author_name": "tomkkk",
      "author_url": "",
      "post_date": "01/31/2021 17:37:49",
      "content": "<p>Article and code that use FP:<br>\n<a href=\"https://github.com/Sieve-Analytics/arbimon2-cnn/blob/master/2_train_ResNet50.ipynb\" target=\"_blank\">https://github.com/Sieve-Analytics/arbimon2-cnn/blob/master/2_train_ResNet50.ipynb</a><br>\n<a href=\"https://www.sciencedirect.com/science/article/pii/S1574954120300637\" target=\"_blank\">https://www.sciencedirect.com/science/article/pii/S1574954120300637</a>  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1175726": "Does anybody have an idea what the False positive dataset and the recordings should be used for ??",
    "1176436": "Some of the labelled TPs have other classes present in the crop, as do the FPs. The only thing we know for *certain* about the FPs is that the particular class is not present. In other words, they are unambiguous data samples for one class only. I've experimented with labels that are 0.5 (or so) for all classes and only 1 or 0 for the class in question, depending on whether or not it's an FP or TP, but it didn't improve my validation procedure. \n\nSomeone cleverer than me has probably found a way to use them to their benefit, but I haven't found them useful.",
    "1176879": "Hi,\n\nYou can use them for data augmentation like background noise. [audiomentations](https://github.com/iver56/audiomentations) library does that for you.",
    "1179637": "Article and code that use FP:\nhttps://github.com/Sieve-Analytics/arbimon2-cnn/blob/master/2_train_ResNet50.ipynb\nhttps://www.sciencedirect.com/science/article/pii/S1574954120300637",
    "1183447": "yes, the challenge is that FP data only gives Not Present information for particular class, and this is a multi-label task. I`m thinking about to play some trick in training loss function, to mask/suppress 23 lables other than the one indicated as 0 in FP data, to let the network only learn from this particular label 0 indicated from FP data. Don`t know if it`s feasible."
  },
  "source": "meta"
}