{
  "id": 191266,
  "title": "Metric ideas and train all labels, maybe can help in the end.",
  "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/191266",
  "author_name": "Kirderf",
  "post_date": "2020-10-15T15:16:33.423000",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Changes the topic to Metric ideas… from about metric sum, saw that \"At the image level, we have a binary classification where the image is classified as PE Present on Image or not (image is negative for PE).\" so it will then sum to 1.0 not 1.07361963, all well.</p>\n<p>I'm tuning the weights in the validation(comp metric) vs training(custom weight) within the 14 labels.<br>\nI have an idea to train all 14 labels and then also try to use the informational labels (one can try adding them to the optimizing to see if they can have an impact) to adjust the logical scheme of the Label Consistency order, if it's necessary, one can hope that the model will sort it all out, best scenario of course.</p>",
  "messages": [
    {
      "id": 1050600,
      "postDate": "2020-10-15T15:16:33.423Z",
      "content": "<p>Changes the topic to Metric ideas… from about metric sum, saw that \"At the image level, we have a binary classification where the image is classified as PE Present on Image or not (image is negative for PE).\" so it will then sum to 1.0 not 1.07361963, all well.</p>\n<p>I'm tuning the weights in the validation(comp metric) vs training(custom weight) within the 14 labels.<br>\nI have an idea to train all 14 labels and then also try to use the informational labels (one can try adding them to the optimizing to see if they can have an impact) to adjust the logical scheme of the Label Consistency order, if it's necessary, one can hope that the model will sort it all out, best scenario of course.</p>",
      "rawMarkdown": "Changes the topic to Metric ideas... from about metric sum, saw that \"At the image level, we have a binary classification where the image is classified as PE Present on Image or not (image is negative for PE).\" so it will then sum to 1.0 not 1.07361963, all well.\n\nI'm tuning the weights in the validation(comp metric) vs training(custom weight) within the 14 labels.\nI have an idea to train all 14 labels and then also try to use the informational labels (one can try adding them to the optimizing to see if they can have an impact) to adjust the logical scheme of the Label Consistency order, if it's necessary, one can hope that the model will sort it all out, best scenario of course.",
      "votes": 1
    },
    {
      "id": 1051911,
      "postDate": "2020-10-17T03:55:15.090Z",
      "content": "<p><a href=\"https://www.kaggle.com/kirderf\" target=\"_blank\">@kirderf</a> does your model converges if making use of metric as loss ?<br>\ni.e calculating pe loss and ct loss</p>",
      "rawMarkdown": "@kirderf does your model converges if making use of metric as loss ?\ni.e calculating pe loss and ct loss",
      "replies": [
        {
          "id": 1052372,
          "postDate": "2020-10-17T16:33:15.403Z",
          "content": "<p>Doing a mix of training right now, one of them are above, and that model I'm using a custom 14 label weight in training and the metric weight in validation. Needed some GPU time, so I'm continue the training now, we'll see the outcome of the testing soon :)</p>",
          "rawMarkdown": "Doing a mix of training right now, one of them are above, and that model I'm using a custom 14 label weight in training and the metric weight in validation. Needed some GPU time, so I'm continue the training now, we'll see the outcome of the testing soon :)"
        }
      ]
    },
    {
      "id": 1050626,
      "postDate": "2020-10-15T15:36:50.130Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1051911,
      "author_name": "Jaideep",
      "author_url": "",
      "post_date": "2020-10-17T03:55:15.090000",
      "content": "<p><a href=\"https://www.kaggle.com/kirderf\" target=\"_blank\">@kirderf</a> does your model converges if making use of metric as loss ?<br>\ni.e calculating pe loss and ct loss</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1052372,
          "author_name": "Kirderf",
          "author_url": "",
          "post_date": "2020-10-17T16:33:15.403000",
          "content": "<p>Doing a mix of training right now, one of them are above, and that model I'm using a custom 14 label weight in training and the metric weight in validation. Needed some GPU time, so I'm continue the training now, we'll see the outcome of the testing soon :)</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1050626,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-10-15T15:36:50.130000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1050600": "Changes the topic to Metric ideas... from about metric sum, saw that \"At the image level, we have a binary classification where the image is classified as PE Present on Image or not (image is negative for PE).\" so it will then sum to 1.0 not 1.07361963, all well.\n\nI'm tuning the weights in the validation(comp metric) vs training(custom weight) within the 14 labels.\nI have an idea to train all 14 labels and then also try to use the informational labels (one can try adding them to the optimizing to see if they can have an impact) to adjust the logical scheme of the Label Consistency order, if it's necessary, one can hope that the model will sort it all out, best scenario of course.",
    "1051911": "@kirderf does your model converges if making use of metric as loss ?\ni.e calculating pe loss and ct loss",
    "1050626": ""
  }
}