{
  "id": 189406,
  "title": "Is this a correct direction",
  "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/189406",
  "author_name": "yuvaramsingh",
  "post_date": "2020-10-07T13:48:30.483000",
  "votes": 1,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi all,<br>\n   I like this competition because of its complex nature of asking image level and study level prediction. currently, i have made an initial level model and made my submission working(after a lot of failed attempts and becoming a Memory leak expert) and got <code>1.285</code> in public score.</p>\n<p>i like to explain my approach so far and it would be really helpful if i can get some comments on where am i going wrong and what should be my next steps to improve my overall model</p>\n<ol>\n<li>Stage 1: Resnet18(for now) based image level feature extractor. This is trained on image level PE label.</li>\n<li>stage 2: The trained Resnet18 is used to extract features from the ct scanes of an exam and passed to Bi-directional LSTM model that is trained on exam level label.</li>\n</ol>\n<p>this model is giving about <code>1.285</code> weighted log loss on leaderboard.</p>\n<p>i would like to improve my current model so it can perform even better. any commands are welcomed. </p>",
  "messages": [
    {
      "id": 1041623,
      "postDate": "2020-10-07T20:56:46.407Z",
      "content": "<p>Not sure why this post is getting down voted . i am just trying to correct my mistake and get better . 👍</p>",
      "rawMarkdown": "Not sure why this post is getting down voted . i am just trying to correct my mistake and get better . 👍",
      "votes": 3
    },
    {
      "id": 1041007,
      "postDate": "2020-10-07T13:48:30.483Z",
      "content": "<p>Hi all,<br>\n   I like this competition because of its complex nature of asking image level and study level prediction. currently, i have made an initial level model and made my submission working(after a lot of failed attempts and becoming a Memory leak expert) and got <code>1.285</code> in public score.</p>\n<p>i like to explain my approach so far and it would be really helpful if i can get some comments on where am i going wrong and what should be my next steps to improve my overall model</p>\n<ol>\n<li>Stage 1: Resnet18(for now) based image level feature extractor. This is trained on image level PE label.</li>\n<li>stage 2: The trained Resnet18 is used to extract features from the ct scanes of an exam and passed to Bi-directional LSTM model that is trained on exam level label.</li>\n</ol>\n<p>this model is giving about <code>1.285</code> weighted log loss on leaderboard.</p>\n<p>i would like to improve my current model so it can perform even better. any commands are welcomed. </p>",
      "rawMarkdown": "Hi all,\n   I like this competition because of its complex nature of asking image level and study level prediction. currently, i have made an initial level model and made my submission working(after a lot of failed attempts and becoming a Memory leak expert) and got `1.285` in public score.\n\ni like to explain my approach so far and it would be really helpful if i can get some comments on where am i going wrong and what should be my next steps to improve my overall model\n\n1. Stage 1: Resnet18(for now) based image level feature extractor. This is trained on image level PE label.\n2. stage 2: The trained Resnet18 is used to extract features from the ct scanes of an exam and passed to Bi-directional LSTM model that is trained on exam level label.\n\nthis model is giving about `1.285` weighted log loss on leaderboard.\n\ni would like to improve my current model so it can perform even better. any commands are welcomed. ",
      "votes": 1
    },
    {
      "id": 1041551,
      "postDate": "2020-10-07T20:04:53.053Z",
      "content": "<p>If you're getting 1.285 in public LB with a model, that's suggesting you're doing something wrong with your inference. A submission of all 0.5 gets 0.693, and the average gets 0.434. </p>",
      "rawMarkdown": "If you're getting 1.285 in public LB with a model, that's suggesting you're doing something wrong with your inference. A submission of all 0.5 gets 0.693, and the average gets 0.434. ",
      "votes": 1,
      "replies": [
        {
          "id": 1041586,
          "postDate": "2020-10-07T20:34:20.503Z",
          "content": "<p>i guess image level prediction have more influence on the overall loss. i am setting the image level prediction to the sigmoid output of the model. do you see anything that needed to be changed in this approach. <br>\nis it because my model is having vary bad prediction. any suggestion on how i can correct this.</p>",
          "rawMarkdown": "i guess image level prediction have more influence on the overall loss. i am setting the image level prediction to the sigmoid output of the model. do you see anything that needed to be changed in this approach. \nis it because my model is having vary bad prediction. any suggestion on how i can correct this."
        },
        {
          "id": 1041589,
          "postDate": "2020-10-07T20:36:01.720Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1041619,
          "postDate": "2020-10-07T20:55:06.273Z",
          "content": "<p>got it . Actually i am now focused on correcting my mistakes on the image level predictions first as of now. i am making a 2 stage network where stage 1 handles image level prediciton (1) and stage 2 takes care of the rest(9). <br>\nwhen i tested my weighted image level log loss locally following the evaluation metrics and its implementation kernels, i found the image level prediction loss is about ~1.29 . this is the reason why i was talking about how am i creating image level  my submission in my precious cmd.<br>\nDo you think am in making any mistake by just copying sigmoid output into image level prediction.</p>",
          "rawMarkdown": "got it . Actually i am now focused on correcting my mistakes on the image level predictions first as of now. i am making a 2 stage network where stage 1 handles image level prediciton (1) and stage 2 takes care of the rest(9). \nwhen i tested my weighted image level log loss locally following the evaluation metrics and its implementation kernels, i found the image level prediction loss is about ~1.29 . this is the reason why i was talking about how am i creating image level  my submission in my precious cmd.\nDo you think am in making any mistake by just copying sigmoid output into image level prediction."
        },
        {
          "id": 1059913,
          "postDate": "2020-10-25T15:25:15.910Z",
          "content": "<p><a href=\"https://www.kaggle.com/stanleyjzheng\" target=\"_blank\">@stanleyjzheng</a> <br>\nExam loss mentioned in competition is weighted sum , when we take mean of it then  should it be for a   sum of weighted loss / number of elements in batch ( which is what i think pytorch's bce with logits does ,when reduction='mean' )   </p>\n<p>or </p>\n<p>should it be sum (weighted loss) of all labels in exam / number of exams in batch ..</p>\n<p>When i do first one i get too less Exam loss.. is that right mean..</p>",
          "rawMarkdown": "@stanleyjzheng \nExam loss mentioned in competition is weighted sum , when we take mean of it then  should it be for a   sum of weighted loss / number of elements in batch ( which is what i think pytorch's bce with logits does ,when reduction='mean' )   \n\n or \n\nshould it be sum (weighted loss) of all labels in exam / number of exams in batch ..\n \n\nWhen i do first one i get too less Exam loss.. is that right mean..\n",
          "votes": -1
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1041623,
      "author_name": "yuvaramsingh",
      "author_url": "",
      "post_date": "2020-10-07T20:56:46.407000",
      "content": "<p>Not sure why this post is getting down voted . i am just trying to correct my mistake and get better . 👍</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 1041551,
      "author_name": "Stanley Zheng",
      "author_url": "",
      "post_date": "2020-10-07T20:04:53.053000",
      "content": "<p>If you're getting 1.285 in public LB with a model, that's suggesting you're doing something wrong with your inference. A submission of all 0.5 gets 0.693, and the average gets 0.434. </p>",
      "votes": 1,
      "replies": [
        {
          "id": 1041586,
          "author_name": "yuvaramsingh",
          "author_url": "",
          "post_date": "2020-10-07T20:34:20.503000",
          "content": "<p>i guess image level prediction have more influence on the overall loss. i am setting the image level prediction to the sigmoid output of the model. do you see anything that needed to be changed in this approach. <br>\nis it because my model is having vary bad prediction. any suggestion on how i can correct this.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1041589,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-10-07T20:36:01.720000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1041619,
          "author_name": "yuvaramsingh",
          "author_url": "",
          "post_date": "2020-10-07T20:55:06.273000",
          "content": "<p>got it . Actually i am now focused on correcting my mistakes on the image level predictions first as of now. i am making a 2 stage network where stage 1 handles image level prediciton (1) and stage 2 takes care of the rest(9). <br>\nwhen i tested my weighted image level log loss locally following the evaluation metrics and its implementation kernels, i found the image level prediction loss is about ~1.29 . this is the reason why i was talking about how am i creating image level  my submission in my precious cmd.<br>\nDo you think am in making any mistake by just copying sigmoid output into image level prediction.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1059913,
          "author_name": "Jaideep",
          "author_url": "",
          "post_date": "2020-10-25T15:25:15.910000",
          "content": "<p><a href=\"https://www.kaggle.com/stanleyjzheng\" target=\"_blank\">@stanleyjzheng</a> <br>\nExam loss mentioned in competition is weighted sum , when we take mean of it then  should it be for a   sum of weighted loss / number of elements in batch ( which is what i think pytorch's bce with logits does ,when reduction='mean' )   </p>\n<p>or </p>\n<p>should it be sum (weighted loss) of all labels in exam / number of exams in batch ..</p>\n<p>When i do first one i get too less Exam loss.. is that right mean..</p>",
          "votes": -1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1041623": "Not sure why this post is getting down voted . i am just trying to correct my mistake and get better . 👍",
    "1041007": "Hi all,\n   I like this competition because of its complex nature of asking image level and study level prediction. currently, i have made an initial level model and made my submission working(after a lot of failed attempts and becoming a Memory leak expert) and got `1.285` in public score.\n\ni like to explain my approach so far and it would be really helpful if i can get some comments on where am i going wrong and what should be my next steps to improve my overall model\n\n1. Stage 1: Resnet18(for now) based image level feature extractor. This is trained on image level PE label.\n2. stage 2: The trained Resnet18 is used to extract features from the ct scanes of an exam and passed to Bi-directional LSTM model that is trained on exam level label.\n\nthis model is giving about `1.285` weighted log loss on leaderboard.\n\ni would like to improve my current model so it can perform even better. any commands are welcomed. ",
    "1041551": "If you're getting 1.285 in public LB with a model, that's suggesting you're doing something wrong with your inference. A submission of all 0.5 gets 0.693, and the average gets 0.434. "
  }
}