{
  "id": 199489,
  "title": "competitive with kaggle resources only",
  "url": "/competitions/riiid-test-answer-prediction/discussion/199489",
  "author_name": "",
  "post_date": "2020-11-25T23:50:01.785019Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>hi</p>\n<p>is it possible to be competitive (say, top 1%) in this competition,<br>\nonly using kaggle's kernel resources (CPU, GPU, memory etc)?<br>\nmeaning, without any local machine power?</p>\n<p>because the data is ~5gb, I wonder if it is possible…<br>\nwould love to hear from the experienced guys here</p>\n<p>thanks</p>",
  "messages": [
    {
      "id": "1091319",
      "postDate": "11/25/2020 23:50:01",
      "content": "<p>hi</p>\n<p>is it possible to be competitive (say, top 1%) in this competition,<br>\nonly using kaggle's kernel resources (CPU, GPU, memory etc)?<br>\nmeaning, without any local machine power?</p>\n<p>because the data is ~5gb, I wonder if it is possible…<br>\nwould love to hear from the experienced guys here</p>\n<p>thanks</p>",
      "rawMarkdown": "hi\n\nis it possible to be competitive (say, top 1%) in this competition,\nonly using kaggle's kernel resources (CPU, GPU, memory etc)?\nmeaning, without any local machine power?\n\nbecause the data is ~5gb, I wonder if it is possible...\nwould love to hear from the experienced guys here\n\nthanks",
      "votes": null
    },
    {
      "id": "1091700",
      "postDate": "11/26/2020 08:05:31",
      "content": "<blockquote>\n  <p>is it possible to be competitive (say, top 1%) in this competition,<br>\n  only using kaggle's kernel resources</p>\n</blockquote>\n<p>I doubt so. People are reporting &gt; 0.01 when going from 10M -&gt; full dataset, and I think there will be more than 1% of people who will be able to use larger resources. Of course you can always catch up with super smart features…</p>",
      "rawMarkdown": "> is it possible to be competitive (say, top 1%) in this competition,\nonly using kaggle's kernel resources\n\nI doubt so. People are reporting > 0.01 when going from 10M -> full dataset, and I think there will be more than 1% of people who will be able to use larger resources. Of course you can always catch up with super smart features...",
      "votes": null
    },
    {
      "id": "1092092",
      "postDate": "11/26/2020 14:33:45",
      "content": "<p>I spent a lot of time on a lot of competitions on Kaggle in past two years.<br>\nFor many of those, I think it would have been possible to stay in the top 10% using only kaggle resources.  </p>\n<p>Unless your very good and per <a href=\"url\" target=\"_blank\">bluetrain</a> you have super smart features I don't think you can get to top 1%.  I believe that most of the very good kagglers also have very good resources beyond Kaggle that they can use.</p>\n<p>For me, the major issue with kernels is they are just not fun with the current limitations.  The limitations of the kernels require that for most competitions, you are forced to a multiple kernels/scripts pipeline to stay in the single submission requirements.  It is a great learning environment to do this kind of work in the kernels, but it's a hard environment to get gold.</p>\n<p>So - get good using the free resources.  When you get your 2nd or 3rd top 10% metal, than explore additional compute.</p>",
      "rawMarkdown": "I spent a lot of time on a lot of competitions on Kaggle in past two years.\nFor many of those, I think it would have been possible to stay in the top 10% using only kaggle resources.  \n\nUnless your very good and per [bluetrain](url) you have super smart features I don't think you can get to top 1%.  I believe that most of the very good kagglers also have very good resources beyond Kaggle that they can use.\n\nFor me, the major issue with kernels is they are just not fun with the current limitations.  The limitations of the kernels require that for most competitions, you are forced to a multiple kernels/scripts pipeline to stay in the single submission requirements.  It is a great learning environment to do this kind of work in the kernels, but it's a hard environment to get gold.\n\nSo - get good using the free resources.  When you get your 2nd or 3rd top 10% metal, than explore additional compute.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1091700,
      "author_name": "stecasasso",
      "author_url": "",
      "post_date": "11/26/2020 08:05:31",
      "content": "<blockquote>\n  <p>is it possible to be competitive (say, top 1%) in this competition,<br>\n  only using kaggle's kernel resources</p>\n</blockquote>\n<p>I doubt so. People are reporting &gt; 0.01 when going from 10M -&gt; full dataset, and I think there will be more than 1% of people who will be able to use larger resources. Of course you can always catch up with super smart features…</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1092092,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "11/26/2020 14:33:45",
      "content": "<p>I spent a lot of time on a lot of competitions on Kaggle in past two years.<br>\nFor many of those, I think it would have been possible to stay in the top 10% using only kaggle resources.  </p>\n<p>Unless your very good and per <a href=\"url\" target=\"_blank\">bluetrain</a> you have super smart features I don't think you can get to top 1%.  I believe that most of the very good kagglers also have very good resources beyond Kaggle that they can use.</p>\n<p>For me, the major issue with kernels is they are just not fun with the current limitations.  The limitations of the kernels require that for most competitions, you are forced to a multiple kernels/scripts pipeline to stay in the single submission requirements.  It is a great learning environment to do this kind of work in the kernels, but it's a hard environment to get gold.</p>\n<p>So - get good using the free resources.  When you get your 2nd or 3rd top 10% metal, than explore additional compute.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1091319": "hi\n\nis it possible to be competitive (say, top 1%) in this competition,\nonly using kaggle's kernel resources (CPU, GPU, memory etc)?\nmeaning, without any local machine power?\n\nbecause the data is ~5gb, I wonder if it is possible...\nwould love to hear from the experienced guys here\n\nthanks",
    "1091700": "> is it possible to be competitive (say, top 1%) in this competition,\nonly using kaggle's kernel resources\n\nI doubt so. People are reporting > 0.01 when going from 10M -> full dataset, and I think there will be more than 1% of people who will be able to use larger resources. Of course you can always catch up with super smart features...",
    "1092092": "I spent a lot of time on a lot of competitions on Kaggle in past two years.\nFor many of those, I think it would have been possible to stay in the top 10% using only kaggle resources.  \n\nUnless your very good and per [bluetrain](url) you have super smart features I don't think you can get to top 1%.  I believe that most of the very good kagglers also have very good resources beyond Kaggle that they can use.\n\nFor me, the major issue with kernels is they are just not fun with the current limitations.  The limitations of the kernels require that for most competitions, you are forced to a multiple kernels/scripts pipeline to stay in the single submission requirements.  It is a great learning environment to do this kind of work in the kernels, but it's a hard environment to get gold.\n\nSo - get good using the free resources.  When you get your 2nd or 3rd top 10% metal, than explore additional compute."
  },
  "source": "meta"
}