{
  "id": 55338,
  "title": "Any success on using LightGBM or XGboost with GPU acceleration?",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/55338",
  "author_name": "",
  "post_date": "2018-04-25T11:10:44.766170900Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "",
  "messages": [
    {
      "id": "319149",
      "postDate": "04/25/2018 11:10:44",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "320216",
      "postDate": "04/27/2018 21:57:26",
      "content": "<p>What is your def of success? \nWe are using LightGBM with gpu and our best single model so far is 0.9789 - is that a success? I don't know. \nAs far as it is our second competition I define success as gaining knowledge ;)\nCheers</p>",
      "rawMarkdown": "What is your def of success? \nWe are using LightGBM with gpu and our best single model so far is 0.9789 - is that a success? I don't know. \nAs far as it is our second competition I define success as gaining knowledge ;)\nCheers",
      "votes": null
    },
    {
      "id": "321250",
      "postDate": "04/30/2018 22:02:20",
      "content": "<p>I was able to setup both XGboost and LightGBM to run using GPU.  I simply followed the installation instructions detailed on their respective websites. Compiled from source and installed python package.\nAre you running into any particular issue with the two packages?</p>",
      "rawMarkdown": "I was able to setup both XGboost and LightGBM to run using GPU.  I simply followed the installation instructions detailed on their respective websites. Compiled from source and installed python package.\nAre you running into any particular issue with the two packages?",
      "votes": null
    },
    {
      "id": "321269",
      "postDate": "04/30/2018 22:31:01",
      "content": "<p>From a performance perspective, has not been successful. Training on GPU seems to be marginally slower than on CPU (1080TI vs i7 8770K) here.</p>",
      "rawMarkdown": "From a performance perspective, has not been successful. Training on GPU seems to be marginally slower than on CPU (1080TI vs i7 8770K) here.",
      "votes": null
    },
    {
      "id": "322106",
      "postDate": "05/02/2018 12:28:22",
      "content": "<p>I was not able to install LightGBM for GPU on my local machine. I have tried with pip but got errors regarding OpenCL. Also tried installing from sources and it broke as well. I have searched for solutions in threads, but still no solution. If you have any links you want to share please do </p>",
      "rawMarkdown": "I was not able to install LightGBM for GPU on my local machine. I have tried with pip but got errors regarding OpenCL. Also tried installing from sources and it broke as well. I have searched for solutions in threads, but still no solution. If you have any links you want to share please do",
      "votes": null
    },
    {
      "id": "322165",
      "postDate": "05/02/2018 13:50:55",
      "content": "<p>Hi Marios. I didn't do anything special -- just ran instructions <a href=\"https://github.com/Microsoft/LightGBM/blob/master/docs/GPU-Tutorial.rst#build-lightgbm\">from here</a> on ubuntu desktop. Then set the appropriate flag <code>'device':'gpu'</code> to pass along to <code>lgb.train</code>. <a href=\"https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/52805#311950\">I've noticed</a> that <code>nthreads</code> may also interact with this, even though gpu training is on. At the time of that post, I was on a different cpu/gpu architecture than now though. I would not recommend spending time attempting to take your lgbm -&gt; gpu at this juncture for this particular contest, as whatever speed gain you'd experience would likely be negligible or inexistent depending on your system. But if you're just doing it so that you can train on gpu generally, then power on =)</p>",
      "rawMarkdown": "Hi Marios. I didn't do anything special -- just ran instructions [from here][1] on ubuntu desktop. Then set the appropriate flag `'device':'gpu'` to pass along to `lgb.train`. [I've noticed][2] that `nthreads` may also interact with this, even though gpu training is on. At the time of that post, I was on a different cpu/gpu architecture than now though. I would not recommend spending time attempting to take your lgbm -&gt; gpu at this juncture for this particular contest, as whatever speed gain you'd experience would likely be negligible or inexistent depending on your system. But if you're just doing it so that you can train on gpu generally, then power on =)\n\n\n  [1]: https://github.com/Microsoft/LightGBM/blob/master/docs/GPU-Tutorial.rst#build-lightgbm\n  [2]: https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/52805#311950",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 320216,
      "author_name": "meykds",
      "author_url": "",
      "post_date": "04/27/2018 21:57:26",
      "content": "<p>What is your def of success? \nWe are using LightGBM with gpu and our best single model so far is 0.9789 - is that a success? I don't know. \nAs far as it is our second competition I define success as gaining knowledge ;)\nCheers</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 321250,
      "author_name": "sauravrt",
      "author_url": "",
      "post_date": "04/30/2018 22:02:20",
      "content": "<p>I was able to setup both XGboost and LightGBM to run using GPU.  I simply followed the installation instructions detailed on their respective websites. Compiled from source and installed python package.\nAre you running into any particular issue with the two packages?</p>",
      "votes": null,
      "replies": [
        {
          "id": 321269,
          "author_name": "authman",
          "author_url": "",
          "post_date": "04/30/2018 22:31:01",
          "content": "<p>From a performance perspective, has not been successful. Training on GPU seems to be marginally slower than on CPU (1080TI vs i7 8770K) here.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 322106,
          "author_name": "pejibaye",
          "author_url": "",
          "post_date": "05/02/2018 12:28:22",
          "content": "<p>I was not able to install LightGBM for GPU on my local machine. I have tried with pip but got errors regarding OpenCL. Also tried installing from sources and it broke as well. I have searched for solutions in threads, but still no solution. If you have any links you want to share please do </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 322165,
          "author_name": "authman",
          "author_url": "",
          "post_date": "05/02/2018 13:50:55",
          "content": "<p>Hi Marios. I didn't do anything special -- just ran instructions <a href=\"https://github.com/Microsoft/LightGBM/blob/master/docs/GPU-Tutorial.rst#build-lightgbm\">from here</a> on ubuntu desktop. Then set the appropriate flag <code>'device':'gpu'</code> to pass along to <code>lgb.train</code>. <a href=\"https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/52805#311950\">I've noticed</a> that <code>nthreads</code> may also interact with this, even though gpu training is on. At the time of that post, I was on a different cpu/gpu architecture than now though. I would not recommend spending time attempting to take your lgbm -&gt; gpu at this juncture for this particular contest, as whatever speed gain you'd experience would likely be negligible or inexistent depending on your system. But if you're just doing it so that you can train on gpu generally, then power on =)</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "319149": "",
    "320216": "What is your def of success? \nWe are using LightGBM with gpu and our best single model so far is 0.9789 - is that a success? I don't know. \nAs far as it is our second competition I define success as gaining knowledge ;)\nCheers",
    "321250": "I was able to setup both XGboost and LightGBM to run using GPU.  I simply followed the installation instructions detailed on their respective websites. Compiled from source and installed python package.\nAre you running into any particular issue with the two packages?",
    "321269": "From a performance perspective, has not been successful. Training on GPU seems to be marginally slower than on CPU (1080TI vs i7 8770K) here.",
    "322106": "I was not able to install LightGBM for GPU on my local machine. I have tried with pip but got errors regarding OpenCL. Also tried installing from sources and it broke as well. I have searched for solutions in threads, but still no solution. If you have any links you want to share please do",
    "322165": "Hi Marios. I didn't do anything special -- just ran instructions [from here][1] on ubuntu desktop. Then set the appropriate flag `'device':'gpu'` to pass along to `lgb.train`. [I've noticed][2] that `nthreads` may also interact with this, even though gpu training is on. At the time of that post, I was on a different cpu/gpu architecture than now though. I would not recommend spending time attempting to take your lgbm -&gt; gpu at this juncture for this particular contest, as whatever speed gain you'd experience would likely be negligible or inexistent depending on your system. But if you're just doing it so that you can train on gpu generally, then power on =)\n\n\n  [1]: https://github.com/Microsoft/LightGBM/blob/master/docs/GPU-Tutorial.rst#build-lightgbm\n  [2]: https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/52805#311950"
  },
  "source": "meta"
}