{
  "id": 17049,
  "title": "R nnet() - how much RAM needed?",
  "url": "/competitions/digit-recognizer/discussion/17049",
  "author_name": "Mark C",
  "post_date": "2015-10-17T14:12:59.133000",
  "votes": 0,
  "comment_count": 3,
  "views": null,
  "content": "<p>Hi,</p>\n\n<p>I'm struggling to use R nnet() neural network package. I keep getting out of memory issues. </p>\n\n<p>My setup is Windows 7 x64, 16GB RAM, running Revolution R 3.2.2 &amp; RStudio.</p>\n\n<p>My training data is 33600 rows by 709 variables. </p>\n\n<p>The issue is the hidden layer size parameter. Can anyone please tell me what sizes they are managing and how much RAM they have on their system? I've not even managed 100 neurons in the hidden layer.</p>\n\n<p>Thanks,</p>\n\n<p>Mark</p>",
  "messages": [
    {
      "id": 99093,
      "postDate": "2015-11-19T11:04:45.993Z",
      "content": "<p>From my experience using nnet or neuralnet is useless for this competition. Try h2o package which is great. <a href=\"https://www.kaggle.com/kobakhit/digit-recognizer/digital-recognizer-in-r\">Here</a> is a working example.</p>",
      "rawMarkdown": "From my experience using nnet or neuralnet is useless for this competition. Try h2o package which is great. [Here][1] is a working example.\r\n\r\n\r\n  [1]: https://www.kaggle.com/kobakhit/digit-recognizer/digital-recognizer-in-r",
      "votes": 3
    },
    {
      "id": 124840,
      "postDate": "2016-06-22T17:53:28.207Z",
      "content": "<p>Hello, \nI have Lubuntu 16.04 and 24Gb of RAM.\nI've tried to use neuralnet and watched RAM used on my computer during computations. Sometimes it would be up to 19.5 Gb. (Still no useful result, but this is another matter.)\nIf you have Windows then be aware that your OS can have restrictions on RAM it uses, in spite of what is installed. For example, &quot;Windows 7&quot; version is limited to use 4 Gb of RAM. With &quot;Windows 8&quot; you may have more, depending on &quot;32bit&quot; or &quot;64bit&quot; or &quot;pro&quot; options. I believe 16Gb is the max. There are higher bounds for Windows 10. They have complete info on their site. </p>",
      "rawMarkdown": "Hello, \r\nI have Lubuntu 16.04 and 24Gb of RAM.\r\nI've tried to use neuralnet and watched RAM used on my computer during computations. Sometimes it would be up to 19.5 Gb. (Still no useful result, but this is another matter.)\r\nIf you have Windows then be aware that your OS can have restrictions on RAM it uses, in spite of what is installed. For example, \"Windows 7\" version is limited to use 4 Gb of RAM. With \"Windows 8\" you may have more, depending on \"32bit\" or \"64bit\" or \"pro\" options. I believe 16Gb is the max. There are higher bounds for Windows 10. They have complete info on their site. "
    },
    {
      "id": 99097,
      "postDate": "2015-11-19T11:38:09.243Z",
      "content": "<p>Thanks Koba. Great write up. I will look at h2o.</p>\n\n<p>My nnet() and neuralnet() runs lasted hours and yielded poor accuracy (~93%). I'm perhaps not using them correctly.</p>",
      "rawMarkdown": "Thanks Koba. Great write up. I will look at h2o.\r\n\r\nMy nnet() and neuralnet() runs lasted hours and yielded poor accuracy (~93%). I'm perhaps not using them correctly."
    },
    {
      "id": 96499,
      "postDate": "2015-10-17T14:12:59.133Z",
      "content": "<p>Hi,</p>\n\n<p>I'm struggling to use R nnet() neural network package. I keep getting out of memory issues. </p>\n\n<p>My setup is Windows 7 x64, 16GB RAM, running Revolution R 3.2.2 &amp; RStudio.</p>\n\n<p>My training data is 33600 rows by 709 variables. </p>\n\n<p>The issue is the hidden layer size parameter. Can anyone please tell me what sizes they are managing and how much RAM they have on their system? I've not even managed 100 neurons in the hidden layer.</p>\n\n<p>Thanks,</p>\n\n<p>Mark</p>",
      "rawMarkdown": "Hi,\r\n\r\nI'm struggling to use R nnet() neural network package. I keep getting out of memory issues. \r\n\r\nMy setup is Windows 7 x64, 16GB RAM, running Revolution R 3.2.2 & RStudio.\r\n\r\nMy training data is 33600 rows by 709 variables. \r\n\r\nThe issue is the hidden layer size parameter. Can anyone please tell me what sizes they are managing and how much RAM they have on their system? I've not even managed 100 neurons in the hidden layer.\r\n\r\nThanks,\r\n\r\nMark"
    }
  ],
  "comments": [
    {
      "id": 99093,
      "author_name": "Koba",
      "author_url": "",
      "post_date": "2015-11-19T11:04:45.993000",
      "content": "<p>From my experience using nnet or neuralnet is useless for this competition. Try h2o package which is great. <a href=\"https://www.kaggle.com/kobakhit/digit-recognizer/digital-recognizer-in-r\">Here</a> is a working example.</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 124840,
      "author_name": "Maiia Bakhova",
      "author_url": "",
      "post_date": "2016-06-22T17:53:28.207000",
      "content": "<p>Hello, \nI have Lubuntu 16.04 and 24Gb of RAM.\nI've tried to use neuralnet and watched RAM used on my computer during computations. Sometimes it would be up to 19.5 Gb. (Still no useful result, but this is another matter.)\nIf you have Windows then be aware that your OS can have restrictions on RAM it uses, in spite of what is installed. For example, &quot;Windows 7&quot; version is limited to use 4 Gb of RAM. With &quot;Windows 8&quot; you may have more, depending on &quot;32bit&quot; or &quot;64bit&quot; or &quot;pro&quot; options. I believe 16Gb is the max. There are higher bounds for Windows 10. They have complete info on their site. </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 99097,
      "author_name": "Mark C",
      "author_url": "",
      "post_date": "2015-11-19T11:38:09.243000",
      "content": "<p>Thanks Koba. Great write up. I will look at h2o.</p>\n\n<p>My nnet() and neuralnet() runs lasted hours and yielded poor accuracy (~93%). I'm perhaps not using them correctly.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "99093": "From my experience using nnet or neuralnet is useless for this competition. Try h2o package which is great. [Here][1] is a working example.\r\n\r\n\r\n  [1]: https://www.kaggle.com/kobakhit/digit-recognizer/digital-recognizer-in-r",
    "124840": "Hello, \r\nI have Lubuntu 16.04 and 24Gb of RAM.\r\nI've tried to use neuralnet and watched RAM used on my computer during computations. Sometimes it would be up to 19.5 Gb. (Still no useful result, but this is another matter.)\r\nIf you have Windows then be aware that your OS can have restrictions on RAM it uses, in spite of what is installed. For example, \"Windows 7\" version is limited to use 4 Gb of RAM. With \"Windows 8\" you may have more, depending on \"32bit\" or \"64bit\" or \"pro\" options. I believe 16Gb is the max. There are higher bounds for Windows 10. They have complete info on their site. ",
    "99097": "Thanks Koba. Great write up. I will look at h2o.\r\n\r\nMy nnet() and neuralnet() runs lasted hours and yielded poor accuracy (~93%). I'm perhaps not using them correctly.",
    "96499": "Hi,\r\n\r\nI'm struggling to use R nnet() neural network package. I keep getting out of memory issues. \r\n\r\nMy setup is Windows 7 x64, 16GB RAM, running Revolution R 3.2.2 & RStudio.\r\n\r\nMy training data is 33600 rows by 709 variables. \r\n\r\nThe issue is the hidden layer size parameter. Can anyone please tell me what sizes they are managing and how much RAM they have on their system? I've not even managed 100 neurons in the hidden layer.\r\n\r\nThanks,\r\n\r\nMark"
  }
}