{
  "id": 346464,
  "title": "One dumb question about training models",
  "url": "/competitions/hubmap-organ-segmentation/discussion/346464",
  "author_name": "",
  "post_date": "2022-08-19T16:08:13.123578400Z",
  "votes": 4,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I have tried to train a swin-based model to do this competition. I'm new to the Kaggle competition. I want to know how to choose the best model from several saved models. Are there any tricks to select a proper model from several saved models (based on val loss, training loss, val dice scores ….)</p>\n<p>This may be a very dumb question but I think this is essential to all works. Thanks 🤒🤒</p>",
  "messages": [
    {
      "id": "1906145",
      "postDate": "08/19/2022 16:08:13",
      "content": "<p>I have tried to train a swin-based model to do this competition. I'm new to the Kaggle competition. I want to know how to choose the best model from several saved models. Are there any tricks to select a proper model from several saved models (based on val loss, training loss, val dice scores ….)</p>\n<p>This may be a very dumb question but I think this is essential to all works. Thanks 🤒🤒</p>",
      "rawMarkdown": "I have tried to train a swin-based model to do this competition. I'm new to the Kaggle competition. I want to know how to choose the best model from several saved models. Are there any tricks to select a proper model from several saved models (based on val loss, training loss, val dice scores ....)\n\nThis may be a very dumb question but I think this is essential to all works. Thanks 🤒🤒",
      "votes": null
    },
    {
      "id": "1906200",
      "postDate": "08/19/2022 17:08:13",
      "content": "<p>the best way to learn is to go the old competition and you have unlimited submission per day (kaggle do not restrict the number of submission once the competition is over).</p>\n<p>just submit all solutions and observe the LB results. (in old competitions, you even see both private and public score)</p>\n<p>then you wold know how to choose in future.</p>\n<p>for this you can try last year kidney ftu segmentation or other old segmentation  competition.</p>",
      "rawMarkdown": "the best way to learn is to go the old competition and you have unlimited submission per day (kaggle do not restrict the number of submission once the competition is over).\n\njust submit all solutions and observe the LB results. (in old competitions, you even see both private and public score)\n\nthen you wold know how to choose in future.\n\nfor this you can try last year kidney ftu segmentation or other old segmentation  competition.",
      "votes": null
    },
    {
      "id": "1906582",
      "postDate": "08/20/2022 02:41:15",
      "content": "<p>Thanks a lot. I will try it.</p>",
      "rawMarkdown": "Thanks a lot. I will try it.",
      "votes": null
    },
    {
      "id": "1906589",
      "postDate": "08/20/2022 02:59:24",
      "content": "<p>after the experiments you may find that there is no sure way to select a model that performs best in public/private LB.</p>\n<p>A better way is to  ensemble/average version of a few \"best potential\" models </p>",
      "rawMarkdown": "after the experiments you may find that there is no sure way to select a model that performs best in public/private LB.\n\nA better way is to  ensemble/average version of a few \"best potential\" models",
      "votes": null
    },
    {
      "id": "1906807",
      "postDate": "08/20/2022 07:57:08",
      "content": "<p>I suggest using an ensemble with varied soft voting ways to perhaps blend your individual models to a single submission, this approach works well for me and I am sure it will work well for you too!</p>",
      "rawMarkdown": "I suggest using an ensemble with varied soft voting ways to perhaps blend your individual models to a single submission, this approach works well for me and I am sure it will work well for you too!",
      "votes": null
    },
    {
      "id": "1906849",
      "postDate": "08/20/2022 08:51:27",
      "content": "<p>ok thanks.  i will try your suggestion to get started with my learning and trial.  🤓</p>",
      "rawMarkdown": "ok thanks.  i will try your suggestion to get started with my learning and trial.  🤓",
      "votes": null
    },
    {
      "id": "1906851",
      "postDate": "08/20/2022 08:52:48",
      "content": "<p>could u please give some examples of 'varied soft voting ways' ?</p>",
      "rawMarkdown": "could u please give some examples of 'varied soft voting ways' ?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1906200,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "08/19/2022 17:08:13",
      "content": "<p>the best way to learn is to go the old competition and you have unlimited submission per day (kaggle do not restrict the number of submission once the competition is over).</p>\n<p>just submit all solutions and observe the LB results. (in old competitions, you even see both private and public score)</p>\n<p>then you wold know how to choose in future.</p>\n<p>for this you can try last year kidney ftu segmentation or other old segmentation  competition.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1906582,
          "author_name": "yingshuli",
          "author_url": "",
          "post_date": "08/20/2022 02:41:15",
          "content": "<p>Thanks a lot. I will try it.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1906589,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "08/20/2022 02:59:24",
          "content": "<p>after the experiments you may find that there is no sure way to select a model that performs best in public/private LB.</p>\n<p>A better way is to  ensemble/average version of a few \"best potential\" models </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1906849,
          "author_name": "yingshuli",
          "author_url": "",
          "post_date": "08/20/2022 08:51:27",
          "content": "<p>ok thanks.  i will try your suggestion to get started with my learning and trial.  🤓</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1906807,
      "author_name": "ravi20076",
      "author_url": "",
      "post_date": "08/20/2022 07:57:08",
      "content": "<p>I suggest using an ensemble with varied soft voting ways to perhaps blend your individual models to a single submission, this approach works well for me and I am sure it will work well for you too!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1906851,
          "author_name": "yingshuli",
          "author_url": "",
          "post_date": "08/20/2022 08:52:48",
          "content": "<p>could u please give some examples of 'varied soft voting ways' ?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1906145": "I have tried to train a swin-based model to do this competition. I'm new to the Kaggle competition. I want to know how to choose the best model from several saved models. Are there any tricks to select a proper model from several saved models (based on val loss, training loss, val dice scores ....)\n\nThis may be a very dumb question but I think this is essential to all works. Thanks 🤒🤒",
    "1906200": "the best way to learn is to go the old competition and you have unlimited submission per day (kaggle do not restrict the number of submission once the competition is over).\n\njust submit all solutions and observe the LB results. (in old competitions, you even see both private and public score)\n\nthen you wold know how to choose in future.\n\nfor this you can try last year kidney ftu segmentation or other old segmentation  competition.",
    "1906582": "Thanks a lot. I will try it.",
    "1906589": "after the experiments you may find that there is no sure way to select a model that performs best in public/private LB.\n\nA better way is to  ensemble/average version of a few \"best potential\" models",
    "1906807": "I suggest using an ensemble with varied soft voting ways to perhaps blend your individual models to a single submission, this approach works well for me and I am sure it will work well for you too!",
    "1906849": "ok thanks.  i will try your suggestion to get started with my learning and trial.  🤓",
    "1906851": "could u please give some examples of 'varied soft voting ways' ?"
  },
  "source": "meta"
}