{
  "id": 115104,
  "title": "A question regarding retraining on stage 2",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/115104",
  "author_name": "Appian",
  "post_date": "2019-10-31T10:43:24.006000",
  "votes": 6,
  "comment_count": 9,
  "views": 0,
  "content": "<p>@juliaelliott</p>\n\n<p>I have a question regarding retraining.</p>\n\n<p>Let's say we have 2 trained models.</p>\n\n<ul>\n<li>model A</li>\n<li>model B</li>\n</ul>\n\n<p>As we are allowed to retrain, we retrain <code>model A</code> using stage 2 dataset and ends up having 3 models at total.</p>\n\n<ul>\n<li>model A</li>\n<li>model B</li>\n<li>model C (model A retrained)</li>\n</ul>\n\n<p>What combinations of emsembling is allowed in this case?\nMy assumtion is <code>B,C</code> is allowed and <code>A,B,C</code> is not allowed as <code>model C</code> is replacing <code>model A</code>. It would be nice if you can clarify.</p>",
  "messages": [
    {
      "id": 662263,
      "postDate": "2019-10-31T10:43:24.007Z",
      "content": "<p>@juliaelliott</p>\n\n<p>I have a question regarding retraining.</p>\n\n<p>Let's say we have 2 trained models.</p>\n\n<ul>\n<li>model A</li>\n<li>model B</li>\n</ul>\n\n<p>As we are allowed to retrain, we retrain <code>model A</code> using stage 2 dataset and ends up having 3 models at total.</p>\n\n<ul>\n<li>model A</li>\n<li>model B</li>\n<li>model C (model A retrained)</li>\n</ul>\n\n<p>What combinations of emsembling is allowed in this case?\nMy assumtion is <code>B,C</code> is allowed and <code>A,B,C</code> is not allowed as <code>model C</code> is replacing <code>model A</code>. It would be nice if you can clarify.</p>",
      "rawMarkdown": "@juliaelliott\n\nI have a question regarding retraining.\n\nLet's say we have 2 trained models.\n\n- model A\n- model B\n\nAs we are allowed to retrain, we retrain `model A` using stage 2 dataset and ends up having 3 models at total.\n\n- model A\n- model B\n- model C (model A retrained)\n\nWhat combinations of emsembling is allowed in this case?\nMy assumtion is `B,C` is allowed and `A,B,C` is not allowed as `model C` is replacing `model A`. It would be nice if you can clarify.\n",
      "votes": 6
    },
    {
      "id": 662282,
      "postDate": "2019-10-31T11:22:47.520Z",
      "content": "<p>as long as you write a readme file on how it is done, then i think it is ok.\nthis read me file will be upload  as well\nfor example:</p>\n\n<p>```\nconsistent approach for  both stage1 and tsage 2:</p>\n\n<ol>\n<li>use dataset xxx to make model A,B,C  </li>\n<li>given a new dataset (with can be stage1 train or  stage2 train = stage1 test), train model D</li>\n<li>apply  ensemble=A+B+C+D on final test test (which can be stage1 test or stage2 test)</li>\n</ol>\n\n<p>```</p>",
      "rawMarkdown": "as long as you write a readme file on how it is done, then i think it is ok.\nthis read me file will be upload  as well\nfor example:\n\n```\nconsistent approach for  both stage1 and tsage 2:\n\n1. use dataset xxx to make model A,B,C  \n2. given a new dataset (with can be stage1 train or  stage2 train = stage1 test), train model D\n3. apply  ensemble=A+B+C+D on final test test (which can be stage1 test or stage2 test)\n\n\n```",
      "votes": 3,
      "replies": [
        {
          "id": 662285,
          "postDate": "2019-10-31T11:32:08.127Z",
          "content": "<p>This sounds reasonable.\nThank you Heng.</p>",
          "rawMarkdown": "This sounds reasonable.\nThank you Heng."
        },
        {
          "id": 662398,
          "postDate": "2019-10-31T14:12:10.857Z",
          "content": "<p>a few things to take note:</p>\n\n<ol>\n<li>your procedure must  produce the csv file you submit for stgae1 and stage2.</li>\n<li>if you do get a prize, you will have a video conference interview to explain you method and software pipleine. kaggle will verify your pipeline and code i think.</li>\n</ol>",
          "rawMarkdown": "a few things to take note:\n\n1.  your procedure must  produce the csv file you submit for stgae1 and stage2.\n2. if you do get a prize, you will have a video conference interview to explain you method and software pipleine. kaggle will verify your pipeline and code i think.",
          "votes": 1
        },
        {
          "id": 662412,
          "postDate": "2019-10-31T14:28:41.580Z",
          "content": "<p>It's a bit of work to ensure everything. Thank you very much.</p>",
          "rawMarkdown": "It's a bit of work to ensure everything. Thank you very much."
        },
        {
          "id": 662873,
          "postDate": "2019-11-01T05:59:09.680Z",
          "content": "<p>Agree, that's what I decide to do (write a readme file). Certainly, I may not have to worry about interview things ; )</p>",
          "rawMarkdown": "Agree, that's what I decide to do (write a readme file). Certainly, I may not have to worry about interview things ; )",
          "votes": 1
        },
        {
          "id": 662876,
          "postDate": "2019-11-01T06:04:11.573Z",
          "content": "<p>Who’s know? There may be great shakeup</p>",
          "rawMarkdown": "Who’s know? There may be great shakeup"
        }
      ]
    },
    {
      "id": 663417,
      "postDate": "2019-11-02T01:32:41.070Z",
      "content": "<p>We are also wondering about this question. Thanks <a href=\"/appian\">@appian</a> and everybody to help clarify this! :D</p>",
      "rawMarkdown": "We are also wondering about this question. Thanks @appian and everybody to help clarify this! :D",
      "votes": 1
    },
    {
      "id": 662287,
      "postDate": "2019-10-31T11:34:57.343Z",
      "content": "<p>Whatever automatic script you can write is OK, it doesn't matter how many models are inside, and it doesn't matter if it depends on the phase of the moon or not (and you mentioned the moon as an external data source in the corresponding discussion thread). The idea is that you just click run and get output for the second stage submission after few hours. So to answer your question, - all combinations are allowed as long as it is automated in code. Hope that clarifies!</p>",
      "rawMarkdown": "Whatever automatic script you can write is OK, it doesn't matter how many models are inside, and it doesn't matter if it depends on the phase of the moon or not (and you mentioned the moon as an external data source in the corresponding discussion thread). The idea is that you just click run and get output for the second stage submission after few hours. So to answer your question, - all combinations are allowed as long as it is automated in code. Hope that clarifies!",
      "votes": 1,
      "replies": [
        {
          "id": 662289,
          "postDate": "2019-10-31T11:37:28.290Z",
          "content": "<p>Yes, it helps. \nThank you!</p>",
          "rawMarkdown": "Yes, it helps. \nThank you!"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 662282,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2019-10-31T11:22:47.520000",
      "content": "<p>as long as you write a readme file on how it is done, then i think it is ok.\nthis read me file will be upload  as well\nfor example:</p>\n\n<p>```\nconsistent approach for  both stage1 and tsage 2:</p>\n\n<ol>\n<li>use dataset xxx to make model A,B,C  </li>\n<li>given a new dataset (with can be stage1 train or  stage2 train = stage1 test), train model D</li>\n<li>apply  ensemble=A+B+C+D on final test test (which can be stage1 test or stage2 test)</li>\n</ol>\n\n<p>```</p>",
      "votes": 3,
      "replies": [
        {
          "id": 662285,
          "author_name": "Appian",
          "author_url": "",
          "post_date": "2019-10-31T11:32:08.127000",
          "content": "<p>This sounds reasonable.\nThank you Heng.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 662398,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2019-10-31T14:12:10.857000",
          "content": "<p>a few things to take note:</p>\n\n<ol>\n<li>your procedure must  produce the csv file you submit for stgae1 and stage2.</li>\n<li>if you do get a prize, you will have a video conference interview to explain you method and software pipleine. kaggle will verify your pipeline and code i think.</li>\n</ol>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 662412,
          "author_name": "Appian",
          "author_url": "",
          "post_date": "2019-10-31T14:28:41.580000",
          "content": "<p>It's a bit of work to ensure everything. Thank you very much.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 662873,
          "author_name": "Yiheng Wang",
          "author_url": "",
          "post_date": "2019-11-01T05:59:09.680000",
          "content": "<p>Agree, that's what I decide to do (write a readme file). Certainly, I may not have to worry about interview things ; )</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 662876,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2019-11-01T06:04:11.573000",
          "content": "<p>Who’s know? There may be great shakeup</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 663417,
      "author_name": "Neuron Engineer",
      "author_url": "",
      "post_date": "2019-11-02T01:32:41.070000",
      "content": "<p>We are also wondering about this question. Thanks <a href=\"/appian\">@appian</a> and everybody to help clarify this! :D</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 662287,
      "author_name": "nosound",
      "author_url": "",
      "post_date": "2019-10-31T11:34:57.343000",
      "content": "<p>Whatever automatic script you can write is OK, it doesn't matter how many models are inside, and it doesn't matter if it depends on the phase of the moon or not (and you mentioned the moon as an external data source in the corresponding discussion thread). The idea is that you just click run and get output for the second stage submission after few hours. So to answer your question, - all combinations are allowed as long as it is automated in code. Hope that clarifies!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 662289,
          "author_name": "Appian",
          "author_url": "",
          "post_date": "2019-10-31T11:37:28.290000",
          "content": "<p>Yes, it helps. \nThank you!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "662263": "@juliaelliott\n\nI have a question regarding retraining.\n\nLet's say we have 2 trained models.\n\n- model A\n- model B\n\nAs we are allowed to retrain, we retrain `model A` using stage 2 dataset and ends up having 3 models at total.\n\n- model A\n- model B\n- model C (model A retrained)\n\nWhat combinations of emsembling is allowed in this case?\nMy assumtion is `B,C` is allowed and `A,B,C` is not allowed as `model C` is replacing `model A`. It would be nice if you can clarify.\n",
    "662282": "as long as you write a readme file on how it is done, then i think it is ok.\nthis read me file will be upload  as well\nfor example:\n\n```\nconsistent approach for  both stage1 and tsage 2:\n\n1. use dataset xxx to make model A,B,C  \n2. given a new dataset (with can be stage1 train or  stage2 train = stage1 test), train model D\n3. apply  ensemble=A+B+C+D on final test test (which can be stage1 test or stage2 test)\n\n\n```",
    "663417": "We are also wondering about this question. Thanks @appian and everybody to help clarify this! :D",
    "662287": "Whatever automatic script you can write is OK, it doesn't matter how many models are inside, and it doesn't matter if it depends on the phase of the moon or not (and you mentioned the moon as an external data source in the corresponding discussion thread). The idea is that you just click run and get output for the second stage submission after few hours. So to answer your question, - all combinations are allowed as long as it is automated in code. Hope that clarifies!"
  }
}