{
  "id": 333724,
  "title": "Emsamblers",
  "url": "/competitions/AI4Code/discussion/333724",
  "author_name": "",
  "post_date": "2022-06-27T23:16:42.983755200Z",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi there all,</p>\n<p>This is my first competition on NLP, and I was wondering if someone can share how to do an ensemble with different models.</p>\n<p>Thanks.</p>",
  "messages": [
    {
      "id": "1835609",
      "postDate": "06/27/2022 23:16:42",
      "content": "<p>Hi there all,</p>\n<p>This is my first competition on NLP, and I was wondering if someone can share how to do an ensemble with different models.</p>\n<p>Thanks.</p>",
      "rawMarkdown": "Hi there all,\n\nThis is my first competition on NLP, and I was wondering if someone can share how to do an ensemble with different models.\n\nThanks.",
      "votes": null
    },
    {
      "id": "1835923",
      "postDate": "06/28/2022 08:07:00",
      "content": "<p>I'm not exactly sure what you mean. But if you mean that you have many models and want to use them all to enhance the overall performance on the problem. You train the models first. Afterwards, you may either:</p>\n<ol>\n<li>Use an statistical method to make a better prediction. Here it depends on the output of your model. If you output the rank, you may use the one with the highest number of predictions. However, if you predict something that has to do with similarity or a probability, you can use average or mean.</li>\n<li>Use another model to predict the final prediction given the predictions from your model. You can use a linear regression (which is explainable and the coefficients may help to find the models having the highest proportions) or a neural network that can learn more difficult representations (non-linear).</li>\n</ol>\n<p>This is as much as I know.</p>\n<p>I hope I got your question right and my answer helps you.</p>",
      "rawMarkdown": "I'm not exactly sure what you mean. But if you mean that you have many models and want to use them all to enhance the overall performance on the problem. You train the models first. Afterwards, you may either:\n\n1.  Use an statistical method to make a better prediction. Here it depends on the output of your model. If you output the rank, you may use the one with the highest number of predictions. However, if you predict something that has to do with similarity or a probability, you can use average or mean.\n1.  Use another model to predict the final prediction given the predictions from your model. You can use a linear regression (which is explainable and the coefficients may help to find the models having the highest proportions) or a neural network that can learn more difficult representations (non-linear).\n\nThis is as much as I know.\n\nI hope I got your question right and my answer helps you.",
      "votes": null
    },
    {
      "id": "1838339",
      "postDate": "06/30/2022 12:28:12",
      "content": "<p>thanks a lot for your answer. I was wondering if you have a sample code for this competition?</p>",
      "rawMarkdown": "thanks a lot for your answer. I was wondering if you have a sample code for this competition?",
      "votes": null
    },
    {
      "id": "1838597",
      "postDate": "06/30/2022 17:06:40",
      "content": "<p>I have 2 publicly available notebooks. One notebook that I used to generate a new dataset. The other is an EDA. I am working on a notebook to generate a model that I will later use for inference. Lately, I did not have time to work on it (especially because it is my first time using hugging face). I hope to at least finish something before the deadline.</p>",
      "rawMarkdown": "I have 2 publicly available notebooks. One notebook that I used to generate a new dataset. The other is an EDA. I am working on a notebook to generate a model that I will later use for inference. Lately, I did not have time to work on it (especially because it is my first time using hugging face). I hope to at least finish something before the deadline.",
      "votes": null
    },
    {
      "id": "1839415",
      "postDate": "07/01/2022 11:44:13",
      "content": "<p>Does it kind like cv competition? Both are DNN models</p>",
      "rawMarkdown": "Does it kind like cv competition? Both are DNN models",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1835923,
      "author_name": "fmakarem",
      "author_url": "",
      "post_date": "06/28/2022 08:07:00",
      "content": "<p>I'm not exactly sure what you mean. But if you mean that you have many models and want to use them all to enhance the overall performance on the problem. You train the models first. Afterwards, you may either:</p>\n<ol>\n<li>Use an statistical method to make a better prediction. Here it depends on the output of your model. If you output the rank, you may use the one with the highest number of predictions. However, if you predict something that has to do with similarity or a probability, you can use average or mean.</li>\n<li>Use another model to predict the final prediction given the predictions from your model. You can use a linear regression (which is explainable and the coefficients may help to find the models having the highest proportions) or a neural network that can learn more difficult representations (non-linear).</li>\n</ol>\n<p>This is as much as I know.</p>\n<p>I hope I got your question right and my answer helps you.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1838339,
          "author_name": "oscarrangel",
          "author_url": "",
          "post_date": "06/30/2022 12:28:12",
          "content": "<p>thanks a lot for your answer. I was wondering if you have a sample code for this competition?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1838597,
          "author_name": "fmakarem",
          "author_url": "",
          "post_date": "06/30/2022 17:06:40",
          "content": "<p>I have 2 publicly available notebooks. One notebook that I used to generate a new dataset. The other is an EDA. I am working on a notebook to generate a model that I will later use for inference. Lately, I did not have time to work on it (especially because it is my first time using hugging face). I hope to at least finish something before the deadline.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1839415,
      "author_name": "tadashi123",
      "author_url": "",
      "post_date": "07/01/2022 11:44:13",
      "content": "<p>Does it kind like cv competition? Both are DNN models</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1835609": "Hi there all,\n\nThis is my first competition on NLP, and I was wondering if someone can share how to do an ensemble with different models.\n\nThanks.",
    "1835923": "I'm not exactly sure what you mean. But if you mean that you have many models and want to use them all to enhance the overall performance on the problem. You train the models first. Afterwards, you may either:\n\n1.  Use an statistical method to make a better prediction. Here it depends on the output of your model. If you output the rank, you may use the one with the highest number of predictions. However, if you predict something that has to do with similarity or a probability, you can use average or mean.\n1.  Use another model to predict the final prediction given the predictions from your model. You can use a linear regression (which is explainable and the coefficients may help to find the models having the highest proportions) or a neural network that can learn more difficult representations (non-linear).\n\nThis is as much as I know.\n\nI hope I got your question right and my answer helps you.",
    "1838339": "thanks a lot for your answer. I was wondering if you have a sample code for this competition?",
    "1838597": "I have 2 publicly available notebooks. One notebook that I used to generate a new dataset. The other is an EDA. I am working on a notebook to generate a model that I will later use for inference. Lately, I did not have time to work on it (especially because it is my first time using hugging face). I hope to at least finish something before the deadline.",
    "1839415": "Does it kind like cv competition? Both are DNN models"
  },
  "source": "meta"
}