{
  "id": 200618,
  "title": "What's \"single model\" exactly?",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/200618",
  "author_name": "",
  "post_date": "2020-12-01T06:52:39.249685100Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi everyone. I'm wondering if there is some definition of \"single model\".</p>\n<ol>\n<li>If I use N-fold CV for one model, is it still called a single model? Note that N-fold CV will train the model N times and combinding the results.</li>\n<li>Is combination of different epochs of one model multiple models or still a single model? <br>\nThank you. Have fun!👍</li>\n</ol>",
  "messages": [
    {
      "id": "1097645",
      "postDate": "12/01/2020 06:52:39",
      "content": "<p>Hi everyone. I'm wondering if there is some definition of \"single model\".</p>\n<ol>\n<li>If I use N-fold CV for one model, is it still called a single model? Note that N-fold CV will train the model N times and combinding the results.</li>\n<li>Is combination of different epochs of one model multiple models or still a single model? <br>\nThank you. Have fun!👍</li>\n</ol>",
      "rawMarkdown": "Hi everyone. I'm wondering if there is some definition of \"single model\".\n1. If I use N-fold CV for one model, is it still called a single model? Note that N-fold CV will train the model N times and combinding the results.\n2. Is combination of different epochs of one model multiple models or still a single model? \nThank you. Have fun!👍",
      "votes": null
    },
    {
      "id": "1097732",
      "postDate": "12/01/2020 07:58:19",
      "content": "<p>I think single model refers to any submission without blending or stacking. In blending, weighted average of multiple model outputs are considered and in stacking a higher level model is trained from multiple model outputs.</p>\n<p>A single model pipeline can still be very complex and sophisticated. So, single model != simple solution.</p>",
      "rawMarkdown": "I think single model refers to any submission without blending or stacking. In blending, weighted average of multiple model outputs are considered and in stacking a higher level model is trained from multiple model outputs.\n\n\n A single model pipeline can still be very complex and sophisticated. So, single model != simple solution.",
      "votes": null
    },
    {
      "id": "1097801",
      "postDate": "12/01/2020 08:49:44",
      "content": "<p>It's unclear what you mean but in general I can say both of the ways you mentioned above are widely used. You can train a single model over kfolds of data or you can train multiple models over kfolds and then during inference take the average of the output to make prediction. Use both and see what gives better results.</p>",
      "rawMarkdown": "It's unclear what you mean but in general I can say both of the ways you mentioned above are widely used. You can train a single model over kfolds of data or you can train multiple models over kfolds and then during inference take the average of the output to make prediction. Use both and see what gives better results.",
      "votes": null
    },
    {
      "id": "1097894",
      "postDate": "12/01/2020 10:21:01",
      "content": "<p>I sometimes see someone saying something like \"achieve *** by a single model\", I want to figure out what it means and how to do comparison with it. </p>",
      "rawMarkdown": "I sometimes see someone saying something like \"achieve *** by a single model\", I want to figure out what it means and how to do comparison with it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1097732,
      "author_name": "keremt",
      "author_url": "",
      "post_date": "12/01/2020 07:58:19",
      "content": "<p>I think single model refers to any submission without blending or stacking. In blending, weighted average of multiple model outputs are considered and in stacking a higher level model is trained from multiple model outputs.</p>\n<p>A single model pipeline can still be very complex and sophisticated. So, single model != simple solution.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1097801,
      "author_name": "koushiksahu",
      "author_url": "",
      "post_date": "12/01/2020 08:49:44",
      "content": "<p>It's unclear what you mean but in general I can say both of the ways you mentioned above are widely used. You can train a single model over kfolds of data or you can train multiple models over kfolds and then during inference take the average of the output to make prediction. Use both and see what gives better results.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1097894,
          "author_name": "dylanliuofficial",
          "author_url": "",
          "post_date": "12/01/2020 10:21:01",
          "content": "<p>I sometimes see someone saying something like \"achieve *** by a single model\", I want to figure out what it means and how to do comparison with it. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1097645": "Hi everyone. I'm wondering if there is some definition of \"single model\".\n1. If I use N-fold CV for one model, is it still called a single model? Note that N-fold CV will train the model N times and combinding the results.\n2. Is combination of different epochs of one model multiple models or still a single model? \nThank you. Have fun!👍",
    "1097732": "I think single model refers to any submission without blending or stacking. In blending, weighted average of multiple model outputs are considered and in stacking a higher level model is trained from multiple model outputs.\n\n\n A single model pipeline can still be very complex and sophisticated. So, single model != simple solution.",
    "1097801": "It's unclear what you mean but in general I can say both of the ways you mentioned above are widely used. You can train a single model over kfolds of data or you can train multiple models over kfolds and then during inference take the average of the output to make prediction. Use both and see what gives better results.",
    "1097894": "I sometimes see someone saying something like \"achieve *** by a single model\", I want to figure out what it means and how to do comparison with it."
  },
  "source": "meta"
}