{
  "id": 12729,
  "title": "Very slow - LDNN",
  "url": "/competitions/datasciencebowl/discussion/12729",
  "author_name": "",
  "post_date": "2015-03-08T17:33:20.490Z",
  "votes": null,
  "comment_count": 4,
  "views": 1222,
  "content": "<p>Hi</p>\n<p>I have been trying to solve this problem using multiple deep learning methods. However each one is extremely slow. Reducing number of filters or iterations significantly reduce the accuracy&nbsp;of the algorithm.&nbsp;</p>\n\n<p>Any inputs on how to increase speed of learning without compromising on number of hidden layers / features / filters / epochs etc? Has anyone been able to utilise the GPU as well as the CPU together in training LDNNs?&nbsp;</p>\n\n<p>What would be the best recommended system configuration for speeding up the learning process? I currently use an iMac with Core i5 2.9 GHZ, 8GB RAM, NVIDIA GTX 660M with 512 MB RAM. This is too slow. I feel I should assemble a system myself, however would performance not be affected without tight integration between the H/W, OS and our code?</p>\n\n<p>Would appreciate thoughts / recommendations on this.</p>\n\n<p>thanks&nbsp;</p>",
  "messages": [
    {
      "id": "65712",
      "postDate": "03/08/2015 17:33:20",
      "content": "<p>Hi</p>\n<p>I have been trying to solve this problem using multiple deep learning methods. However each one is extremely slow. Reducing number of filters or iterations significantly reduce the accuracy&nbsp;of the algorithm.&nbsp;</p>\n\n<p>Any inputs on how to increase speed of learning without compromising on number of hidden layers / features / filters / epochs etc? Has anyone been able to utilise the GPU as well as the CPU together in training LDNNs?&nbsp;</p>\n\n<p>What would be the best recommended system configuration for speeding up the learning process? I currently use an iMac with Core i5 2.9 GHZ, 8GB RAM, NVIDIA GTX 660M with 512 MB RAM. This is too slow. I feel I should assemble a system myself, however would performance not be affected without tight integration between the H/W, OS and our code?</p>\n\n<p>Would appreciate thoughts / recommendations on this.</p>\n\n<p>thanks&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65748",
      "postDate": "03/09/2015 03:43:53",
      "content": "<p>Abhishek, what is your exactly performance look like, i.e., how much time did you cost in one epoch?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65770",
      "postDate": "03/09/2015 11:38:55",
      "content": "<p>Approx 20 minutes for each epoch. Prediction took longer (40 mins per cycle).&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65772",
      "postDate": "03/09/2015 12:30:40",
      "content": "<p>As we all know, such cost depends on your network architecture and number of neurons for each layer. But 20 mins per epoch seem to long. As for your reference, I use a 7-layers CNN and based on a GPU, I get 20 secs per epoch.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "65964",
      "postDate": "03/11/2015 15:41:47",
      "content": "<p>&nbsp;I dont know much about Mac but I can do some comments about what I have done, If you have also integrated GPU then I would switch to weak one so the Nvidia doesnt need to use additional threads for refreshing x server. You can turn off startup applications to give more space to CPU. Check your running applications and turn them off if you dont need. If you think your machine doesnt help anymore maybe its time to consider overclocking bot cpu and gpu even you can combine this system with your new setup (I dont know how Mac components overclock friendly and how risky it is)</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 65748,
      "author_name": "pengpai",
      "author_url": "",
      "post_date": "03/09/2015 03:43:53",
      "content": "<p>Abhishek, what is your exactly performance look like, i.e., how much time did you cost in one epoch?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65770,
      "author_name": "abhisheksimhadri",
      "author_url": "",
      "post_date": "03/09/2015 11:38:55",
      "content": "<p>Approx 20 minutes for each epoch. Prediction took longer (40 mins per cycle).&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65772,
      "author_name": "pengpai",
      "author_url": "",
      "post_date": "03/09/2015 12:30:40",
      "content": "<p>As we all know, such cost depends on your network architecture and number of neurons for each layer. But 20 mins per epoch seem to long. As for your reference, I use a 7-layers CNN and based on a GPU, I get 20 secs per epoch.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 65964,
      "author_name": "hrgiger",
      "author_url": "",
      "post_date": "03/11/2015 15:41:47",
      "content": "<p>&nbsp;I dont know much about Mac but I can do some comments about what I have done, If you have also integrated GPU then I would switch to weak one so the Nvidia doesnt need to use additional threads for refreshing x server. You can turn off startup applications to give more space to CPU. Check your running applications and turn them off if you dont need. If you think your machine doesnt help anymore maybe its time to consider overclocking bot cpu and gpu even you can combine this system with your new setup (I dont know how Mac components overclock friendly and how risky it is)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "65712": "",
    "65748": "",
    "65770": "",
    "65772": "",
    "65964": ""
  },
  "source": "meta"
}