{
  "id": 395729,
  "title": "Do I need to save models for each fold?",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/395729",
  "author_name": "",
  "post_date": "2023-03-18T14:03:55.895636800Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have a somewhat generic framework that I’ve been developing in the Playground series.   For instance:</p>\n<p><a href=\"https://www.kaggle.com/code/mmellinger66/s3e10-pulsar-models\" target=\"_blank\">https://www.kaggle.com/code/mmellinger66/s3e10-pulsar-models</a></p>\n<p>I’m going to start with this kernel for the Vesuvius Challenge.  In my current kernel, I save the oof and target predictions but I don’t save the models that predicted the values.  To improve my predictions, I’m going to try to use other ensembling techniques.</p>\n<p>However, since I didn’t save my fold models, it seems like it won’t be easy to figure out which ratio of weights are best.  Some ensemble methods simply use the predictions without models, but I’d like to understand how the better ensemble models work.  Do I need to save the trained fold models?</p>",
  "messages": [
    {
      "id": "2187219",
      "postDate": "03/18/2023 14:03:55",
      "content": "<p>I have a somewhat generic framework that I’ve been developing in the Playground series.   For instance:</p>\n<p><a href=\"https://www.kaggle.com/code/mmellinger66/s3e10-pulsar-models\" target=\"_blank\">https://www.kaggle.com/code/mmellinger66/s3e10-pulsar-models</a></p>\n<p>I’m going to start with this kernel for the Vesuvius Challenge.  In my current kernel, I save the oof and target predictions but I don’t save the models that predicted the values.  To improve my predictions, I’m going to try to use other ensembling techniques.</p>\n<p>However, since I didn’t save my fold models, it seems like it won’t be easy to figure out which ratio of weights are best.  Some ensemble methods simply use the predictions without models, but I’d like to understand how the better ensemble models work.  Do I need to save the trained fold models?</p>",
      "rawMarkdown": "I have a somewhat generic framework that I’ve been developing in the Playground series.   For instance:\n\nhttps://www.kaggle.com/code/mmellinger66/s3e10-pulsar-models\n\nI’m going to start with this kernel for the Vesuvius Challenge.  In my current kernel, I save the oof and target predictions but I don’t save the models that predicted the values.  To improve my predictions, I’m going to try to use other ensembling techniques.\n\nHowever, since I didn’t save my fold models, it seems like it won’t be easy to figure out which ratio of weights are best.  Some ensemble methods simply use the predictions without models, but I’d like to understand how the better ensemble models work.  Do I need to save the trained fold models?",
      "votes": null
    },
    {
      "id": "2187385",
      "postDate": "03/18/2023 17:26:41",
      "content": "<p>Depending on what ensembling technique you use. The simplest ensembling techniques average over <code>k</code> predictions. For something more sophisticated, such as Stochastic Weight Averaging, average weights during SGD-based training with cyclical or constant learning rate.</p>\n<p>As for your question, it depends on which technique you decide to use, but in general, it's probably a good idea to save the weights for every fold.</p>",
      "rawMarkdown": "Depending on what ensembling technique you use. The simplest ensembling techniques average over `k` predictions. For something more sophisticated, such as Stochastic Weight Averaging, average weights during SGD-based training with cyclical or constant learning rate.\n\nAs for your question, it depends on which technique you decide to use, but in general, it's probably a good idea to save the weights for every fold.",
      "votes": null
    },
    {
      "id": "2212659",
      "postDate": "04/07/2023 00:36:23",
      "content": "<p>Hi Michael,</p>\n<p>Nothing to do with this fantastic Millenial Vesuvius competition (since you close your comments on  Playground Ep. 12 to avoid some birds of prey eager to be GMs):  </p>\n<p>Awesome script and now I know what's  DRY \"Don't repeat yourself\" <br>\nFor women, that's a classic since we hate to repeat even clothes : )<br>\nThank you.<br>\nBest regards,<br>\nMarília.</p>",
      "rawMarkdown": "Hi Michael,\n\nNothing to do with this fantastic Millenial Vesuvius competition (since you close your comments on  Playground Ep. 12 to avoid some birds of prey eager to be GMs):  \n\nAwesome script and now I know what's  DRY \"Don't repeat yourself\" \nFor women, that's a classic since we hate to repeat even clothes : )\nThank you.\nBest regards,\nMarília.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2187385,
      "author_name": "danielhavir",
      "author_url": "",
      "post_date": "03/18/2023 17:26:41",
      "content": "<p>Depending on what ensembling technique you use. The simplest ensembling techniques average over <code>k</code> predictions. For something more sophisticated, such as Stochastic Weight Averaging, average weights during SGD-based training with cyclical or constant learning rate.</p>\n<p>As for your question, it depends on which technique you decide to use, but in general, it's probably a good idea to save the weights for every fold.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2212659,
      "author_name": "mpwolke",
      "author_url": "",
      "post_date": "04/07/2023 00:36:23",
      "content": "<p>Hi Michael,</p>\n<p>Nothing to do with this fantastic Millenial Vesuvius competition (since you close your comments on  Playground Ep. 12 to avoid some birds of prey eager to be GMs):  </p>\n<p>Awesome script and now I know what's  DRY \"Don't repeat yourself\" <br>\nFor women, that's a classic since we hate to repeat even clothes : )<br>\nThank you.<br>\nBest regards,<br>\nMarília.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2187219": "I have a somewhat generic framework that I’ve been developing in the Playground series.   For instance:\n\nhttps://www.kaggle.com/code/mmellinger66/s3e10-pulsar-models\n\nI’m going to start with this kernel for the Vesuvius Challenge.  In my current kernel, I save the oof and target predictions but I don’t save the models that predicted the values.  To improve my predictions, I’m going to try to use other ensembling techniques.\n\nHowever, since I didn’t save my fold models, it seems like it won’t be easy to figure out which ratio of weights are best.  Some ensemble methods simply use the predictions without models, but I’d like to understand how the better ensemble models work.  Do I need to save the trained fold models?",
    "2187385": "Depending on what ensembling technique you use. The simplest ensembling techniques average over `k` predictions. For something more sophisticated, such as Stochastic Weight Averaging, average weights during SGD-based training with cyclical or constant learning rate.\n\nAs for your question, it depends on which technique you decide to use, but in general, it's probably a good idea to save the weights for every fold.",
    "2212659": "Hi Michael,\n\nNothing to do with this fantastic Millenial Vesuvius competition (since you close your comments on  Playground Ep. 12 to avoid some birds of prey eager to be GMs):  \n\nAwesome script and now I know what's  DRY \"Don't repeat yourself\" \nFor women, that's a classic since we hate to repeat even clothes : )\nThank you.\nBest regards,\nMarília."
  },
  "source": "meta"
}